remove-ai-watermarks/docs/synthid-detector-removal-plan.md

SynthID detector and pixel-only removal research plan

Chronological mixed archive of detector, classifier, and removal work. Not a statement of current product capability. Read the split pages first:

• SynthID local detector research
• Classifier models
• SynthID mark removal research

Shipped behavior remains in supported signals, known limitations, and module internals. Keep this file for dated H-gates, corpora, and session notes. Do not add new campaign results here; put them on the matching split page.

Objective

Build two independent provider tracks with four capabilities:

1. a local, metadata-independent detector for the OpenAI SynthID image signal;
2. a local, metadata-independent detector for the Google SynthID image signal;
3. a pixel-only remover for the OpenAI signal that does not use diffusion, VAE reconstruction, semantic regeneration, or generative inpainting;
4. an independently calibrated pixel-only remover for the Google signal under the same constraints.

OpenAI is the first research track because it has a documented remote verifier. Google follows with the same experimental protocol but its own corpus, labels, model, thresholds, and oracle. No carrier, feature, score, or operating point transfers between providers until a held-out experiment demonstrates that it does.

The detector target is signal presence, not payload recovery, provider classification, or general AI-image classification. The removal target is a minimal pixel residual that makes a source-positive image negative in the matching provider oracle while preserving the image's geometry and semantics.

Non-negotiable evidence rules

1. Detector before remover. A remover may be prototyped against synthetic carriers, but no real-image removal claim is made until the local detector passes its own held-out gate.
2. Provider-specific ground truth. OpenAI and Google labels come from their matching verifier. C2PA is recorded separately and is never the pixel label.
3. No metadata or export leakage. Training and evaluation operate on decoded pixels after controlled metadata stripping and matched re-encoding. Geometry, filename, file size, chunks, encoder settings, and source directories cannot be model inputs.
4. Signal identifiability is a gate. A classifier trained only on provider positives and unrelated negatives can learn the provider's generator or export fingerprint. It is called a provider classifier, not a SynthID detector, until at least one causal control succeeds.
5. The oracle is held out from optimization. Candidate algorithms and hyperparameters are selected locally. Oracle batches are immutable and registered before submission. A remote binary verdict is never used as an online loss or hill-climbing signal.
6. A local score decrease is not removal evidence. Removal requires a source-positive, output-negative result from the matching provider oracle.
7. One hypothesis does not become several signals. Correlated spatial, spectral, and color statistics derived from one residual are reported as one line of evidence unless independent controls separate them.
8. Every result is reproducible. Input and output hashes, code revision, model artifact hash, preprocessing, transform lineage, seed, score, threshold, quality metrics, oracle result, session, and timestamp are retained.

Oracle boundary

OpenAI documents a synchronous Content Provenance API at POST /v1/content_provenance_checks. For images it returns separate c2pa and synthid entries and supports PNG, JPEG, and WebP. It remains a remote, OpenAI-scoped verifier, not a released local decoder. The same documentation explicitly says not to use repeated queries to reverse-engineer, remove, or evade a watermark. Adaptive detector or remover research against that endpoint therefore requires explicit OpenAI authorization or a separate research oracle whose terms permit the work. Without that authorization, the OpenAI track may develop local hypotheses but stops before oracle-driven calibration and removal certification.

Google's Gemini verification flow is a Google-scoped oracle. Its documented result can be detected, not detected, or unclear, and the consumer flow has a small rolling quota. An unclear result is indeterminate, never a negative.

For every permitted oracle batch:

• verify the untouched source first;
• submit one file per request;
• record C2PA and SynthID outcomes independently when both are returned;
• preserve the exact submitted bytes and SHA-256 outside the public repository;
• retry only a transient failure on the same bytes;
• record detected, not_detected, indeterminate, or refused verbatim;
• submit a matched transform-only control before attributing a negative result to an experimental edit;
• reserve a final temporal holdout that no feature, threshold, or remover has seen.

Data design

Corpus layers

Each provider gets a separate corpus with five layers:

Layer	Purpose	Required controls
Verified positives	Learn and evaluate the real signal	Matching provider oracle, original bytes
Same-provider hard negatives	Separate watermark from generator identity	Same surface or model, oracle-negative
External hard negatives	Measure false positives	Cameras, scans, edited photos, other generators, synthetic graphics
Low-texture probes	Expose weak shared structure	Solid colors, gradients, ramps, checkerboards, sparse edges
Causal pairs	Attribute a residual to the watermark	Same underlying pixels, positive and confirmed negative

The causal-pair layer is the most valuable and the hardest to obtain. An authorized encoder-off pair is ideal. A provider output and a pixel-only processed version become a usable pair only after the original is positive and the processed bytes are negative in the provider oracle. Public third-party pairs are discovery material until their provenance and both labels are independently verified.

A negative created by a remover trained against detector A cannot train or validate detector A in the same experiment. Reserve it for detector B or a later model epoch after closing the originating experiment. Otherwise detector and remover can certify each other's shared blind spot.

If no same-provider hard negatives or causal pairs can be obtained, learned real-image models stay explicitly labeled as provider classifiers. Spectral repeatability on solid fills alone does not clear this gate.

Strata

Record and split by:

• provider, product surface, model family, and generation date window;
• native width, height, aspect ratio, file format, and color mode;
• photoreal, face, text-heavy, flat graphic, illustration, low texture, and high texture content;
• untouched, metadata-stripped, lossless-normalized, JPEG/WebP, resized, cropped, and color-adjusted lineage;
• generation session and prompt family;
• parent hash for every derivative.

Store research media under .local-eval/synthid/, not in the public repository. Only cleared fixtures may enter data/synthid/. Tracked files contain schemas, scripts, synthetic fixtures, aggregate verdicts, and non-sensitive hashes.

Split discipline

Deduplicate by decoded-pixel hash and perceptual similarity before splitting. Keep every derivative, prompt sibling, and semantic near-duplicate in one hash group. Split by group into train, validation, locked test, and a later temporal test collected after the detector is frozen. A random image-level split is invalid because it leaks transform and generation-family fingerprints.

The negative set must be large enough to support the claimed operating point. At zero observed false positives, roughly 3,000 independent negatives are needed merely to put the one-sided 95% upper bound near 0.1% by the rule of three. The final evaluation should prefer at least 10,000 hard and ordinary negatives per provider, report confidence intervals, and report every negative stratum separately rather than hiding a weak stratum in an aggregate.

Detector program

Experiment D0: oracle and corpus integrity

Goal: prove that labels and bytes mean what the manifest says.

• Separate SynthID from C2PA in every response.
• Confirm that metadata stripping changes C2PA but does not silently define the SynthID label.
• Confirm every positive used for evaluation on its matching oracle.
• Mutation-test manifest ingestion with swapped provider, duplicate bytes, derivative leakage, an indeterminate result mislabeled as negative, and a changed file after hashing.
• Measure permitted same-byte verifier reproducibility on a small preregistered sample. Do not turn retries into adaptive querying.

Gate: no ambiguous label path, no cross-provider oracle substitution, and no train/test family leakage.

Experiment D1: export and generator confounds

Goal: determine how accurately SynthID can appear to be detected when no watermark-specific evidence is available.

Train deliberately confounded baselines on file/container fields, dimensions, RGB thumbnails, and generator-vs-camera content. Then repeat after canonical decode, metadata removal, resolution matching, and hard-negative balancing.

Gate: any proposed signal feature must beat the canonicalized confound baseline on same-provider hard negatives and on a temporal holdout. Otherwise the result is generator attribution.

Experiment D2: low-texture carrier discovery

Goal: test whether a repeatable carrier component is observable when scene texture is suppressed.

Extend the existing scripts/synthid_pixel_probe.py measurements from grayscale NCC to:

• per-channel and opponent-color residuals;
• two-dimensional FFT magnitude and circular phase coherence;
• wavelet bands and multi-scale autocorrelation;
• resolution and aspect-ratio registration;
• cross-color, cross-session, cross-model, and cross-date agreement;
• matched clean synthetic fills and same-provider oracle-negative probes.

Use leave-one-color, leave-one-session, and leave-one-resolution-out tests. Fixed bins discovered and evaluated on the same images are descriptive only.

Gate: a template learned on one subset must detect held-out positive probes above matched negatives and survive a later collection window. Failure kills the fixed-carrier branch but not the content-dependent branch.

Experiment D3: classical real-image detector

Goal: establish the strongest interpretable baseline before a neural model.

Candidate features include locally normalized high-pass residuals, FFT and wavelet energy ratios, circular phase coherence, color-channel agreement, block periodicity, and correlations against provider/resolution templates. Fit regularized logistic regression and a shallow tree model. Calibration uses only validation data.

Report TPR at 0.1% FPR as the primary metric, with bootstrap confidence intervals. AUROC, average precision, and an arbitrary accuracy percentage are secondary. Also report worst-stratum FPR, temporal-holdout TPR, and score drift.

Gate: advance the classical detector only if the locked test shows a stable watermark-specific advantage over confound baselines. Do not choose a threshold from the locked test.

Mechanism hypotheses for D3-D5

The next detector epoch tests these hypotheses as one preregistered program:

1. Counterfactual labels. Train and calibrate on source-matched clean/watermarked examples, ideally the same underlying image before and after encoding. A result that vanishes when source, date, codec, and dimensions are balanced is a provider fingerprint, not watermark evidence.
2. Canonical full-field evidence. Preserve a 512x512 image-level field, absolute residual amplitude, and chroma alongside normalized residuals. Compare it against the frozen patch baseline on source-disjoint pairs. Failure to improve paired low-FPR detection rejects the added global context.
3. Registration. Score several bicubic canonical views spanning small scale, crop, and offset changes, then aggregate before calibration. This tests whether synchronization is distributed across the image rather than fixed to one global phase or local grid.
4. Independent detection head. Learn one presence logit directly. Treat payload-like or phase-consistency heads as auxiliary evidence, not as the presence decision. Their inclusion must improve a held-out paired test, not only attack-surface scores.
5. Symmetric transformation channel. Apply the identical sampled codec, resize, crop, color, noise, or overlay operation to both members of a pair. A transform is admitted only when the reference decoder, where available, confirms that the transformed positive remains valid.
6. Encoder versions. Compare one universal model with version- or epoch-specific experts on a cross-version transfer matrix. A version split is useful only if it improves held-out likelihood without source metadata at inference.
7. Content-dependent watermarkability. Measure flat, low-contrast, monochrome, logo, and pixel-art strata separately. Allow a content-conditioned expert or abstention instead of forcing one global operating point.
8. Two-sided calibration. Calibrate empirical evidence for both not watermarked and watermarked, returning positive, negative, or abstain. External generator corpora enter afterward as an untouched false-positive challenge, not as a substitute for counterfactual negatives.

Experiment D4: learned residual detector

Goal: learn watermark presence from counterfactual image-level evidence that a fixed template misses.

The primary model uses a full canonical field with raw RGB or luminance/chroma, absolute-amplitude fine and coarse residuals, and optional stationary-wavelet and complex-frequency branches. Every optional representation is encoded separately and fused late; early channel concatenation is not a valid ablation. Locally normalized patch evidence remains a frozen baseline, not the primary input. Multi-view registration is aggregated into one image-level presence logit. Payload-like, phase, localization, and content-watermarkability heads are auxiliary and must prove an incremental held-out benefit.

Use group-aware pair sampling and apply every sampled transformation symmetrically to the clean and watermarked members. Never expose metadata, paths, native dimensions without normalization, or encoder-specific byte patterns to the network. If causal pairs remain unavailable, do not advance a learned model from provider classification to this experiment.

Train provider-specific models first. A shared backbone with provider-specific heads is a later ablation, not the default architecture. Keep a second detector family completely outside remover training so it can reveal surrogate overfit. Compare a universal head with version-specific experts, and calibrate the selected model with two-sided empirical evidence so ambiguous inputs abstain.

Gate for a detector release candidate:

• empirical FPR at or below 0.1% on the locked negative set;
• one-sided 95% TPR lower bound at or above 90% on untouched positives;
• no declared hard-negative stratum above 0.5% FPR;
• useful TPR on the temporal holdout and transformation suite;
• meaningful discrimination on causal pairs or same-provider hard negatives;
• calibrated abstain behavior outside supported providers and strata.

These are research gates, not promises that the proprietary decoder's operating point has been reproduced.

Experiment D5: robustness and drift

Evaluate identity, JPEG and WebP, resize, crop, padding, rotation, color changes, blur, noise, overlays, screenshots, and combinations. Preserve a matched transform-only positive control for every attack family. Report both average and worst-transform TPR at the frozen threshold.

Repeat a small fixed collection after provider model or surface changes. A shift in score distribution opens a new model epoch; it does not silently recalibrate the old threshold.

Localization program

A whole-image detector is not automatically a useful removal loss. Test whether its evidence is spatially causal with three independent methods:

• tile occlusion and replacement with distortion-matched controls;
• detector-gradient attribution for the learned model;
• phase or template energy mapped back to spatial blocks.

For each source, create top-ranked, random, and bottom-ranked edits with equal pixel norm and the same codec path. The oracle batch is registered before any results return.

Gate: top-ranked edits must reduce matching-oracle detection more often than random edits under paired analysis. Aggregate per-image differences and confirm the direction with a sign test. If localization does not transfer, do not build a region remover around it.

Pixel-only removal program

All candidates preserve image dimensions and avoid a generative decoder. Run the following ladder in order.

Experiment R1: analytical carrier subtraction

Estimate provider, model-epoch, and geometry-specific residual components from verified probes or causal pairs. Test:

• complex FFT projection with conjugate symmetry;
• wavelet-band projection;
• per-channel and opponent-color residual subtraction;
• spatially varying strength based on local texture and detector attribution.

Sweep signed amplitude, not only attenuation. Include a sham edit with identical pixel norm outside candidate bins. This is the highest-value path because a real shared post-hoc residual could be removed at very high fidelity.

Experiment R2: constrained per-image optimization

Optimize pixels against an ensemble of frozen local detectors. The objective combines detector margin with L-infinity and L2 bounds, LPIPS or DISTS, MS-SSIM, edge consistency, OCR preservation, and face-embedding preservation where applicable. Expectation over transformations covers lossless export, JPEG/WebP, small resize, and color conversion so the result is not a fragile local adversarial example.

The optimization may read only local detectors. The provider oracle evaluates a frozen candidate batch afterward and never supplies gradients, search direction, or per-step feedback.

Experiment R3: feed-forward residual remover

If R2 transfers to the provider oracle, distill successful minimal residuals into a compact image-to-residual network. Constrain the output amplitude and frequency distribution explicitly. Train on one detector ensemble and select on the held-out detector family. Preserve the per-image optimizer as the reference implementation.

Experiment R4: detector-remover co-evaluation

Evaluate four combinations separately:

1. local detector positive, provider oracle positive;
2. local detector negative, provider oracle positive;
3. local detector positive, provider oracle negative;
4. both negative.

Case 2 is the critical surrogate-overfit failure. Add every permitted example to a future hard set only after the current experiment is closed; never tune and report on the same oracle failure.

Removal success gate

A provider-specific remover is a release candidate only when a locked, source-positive evaluation shows all of the following:

• at least 90% matching-oracle not_detected results, with indeterminate counted as failure;
• a positive matched transform-only control for every evaluated source;
• exact-size output with no semantic regeneration stage;
• median PSNR at least 40 dB and fifth-percentile PSNR at least 35 dB;
• median SSIM at least 0.99, plus LPIPS or DISTS reported rather than optimized silently;
• no OCR regression on text strata and no material face-identity regression on face strata under preregistered thresholds;
• no worse oracle-negative rate after a standard downstream JPEG/WebP/resize suite;
• byte-identical pass-through for detector-negative inputs by default;
• a significant paired advantage over random and norm-matched sham edits;
• no claim of forensic cleanliness without a separately trained removal-artifact detector and a held-out evaluation.

If pixel-only methods fail this gate, retain diffusion regeneration as the explicit fallback. Do not combine its success with pixel-only results.

Provider sequence

OpenAI first

1. Resolve the oracle authorization gate.
2. Build the corpus schema and confound challenge.
3. Verify or reject the shared-carrier hypothesis on low-texture probes.
4. Train classical and learned detectors.
5. Freeze the detector and temporal test.
6. Run R1, then R2, then R3 only after each preceding gate passes.
7. Submit one preregistered final oracle batch.

OpenAI work establishes the experimental machinery, not parameters for Google.

Google second

Repeat the full sequence with Google-native positives, Google hard negatives, and the Gemini oracle. Spend the small manual oracle budget on controls and decisive boundary points, not uniform sweeps. Use local detector uncertainty to choose a batch before submission, then freeze it. Test Gemini app and AI Studio surfaces as separate strata because their export and metadata paths differ.

Implementation order

The research harness remains outside the public API until the gates pass.

1. Implemented: use the private corpus schema and auditor documented in data/synthid/research-manifest.md to record provider, surface, model epoch, session, content stratum, parent hash, transform lineage, separate C2PA and SynthID outcomes, oracle session, and artifact hashes.
2. Implemented: build a label-free local inventory before promotion so byte-identical files, decoded-pixel duplicates, and unsupported formats are visible without inferring evidence from directory names.
3. Implemented: add a corpus auditor that rejects hash-group leakage, missing parent links, ambiguous labels, and unsupported oracle-provider pairs.
4. Harness implemented; evidence run pending: run the manifest-driven D1 challenge over container, thumbnail, and canonical decoded-content baselines. Freeze its validation threshold and report same-provider negative cohorts separately.
5. Generalize scripts/synthid_pixel_probe.py into reusable feature extraction while preserving its current synthetic tests.
6. Add reproducible train/evaluate commands whose output is a versioned model card and metrics snapshot, never an unversioned console claim.
7. Add analytical and optimization removal harnesses with norm-matched controls.
8. Reuse the existing fidelity scripts for PSNR, SSIM, OCR, face, and edge measurements, adding only missing metrics.
9. Package provider-specific detector weights behind an optional dependency only after the detector gate passes.
10. Add a runtime remover and CLI surface only after the removal gate passes.

Pure feature, manifest, split, threshold, and residual-constraint logic must be unit-tested without model downloads. Real model and oracle runs stay explicit research jobs.

The inventory, manifest auditor, and D1 confound harness now exist as local research tools. D1 has not produced a real-corpus metric yet because existing artifacts have not been promoted into evidence-bearing provider manifests. This is an evidence gap, not permission to infer labels from their paths. The next real D1 run begins only after ordinary rows cover train, validation, locked test, same-provider hard negatives, and a two-class temporal holdout.

Hypothesis register

What is open, what is closed, and what would move each one. A status changes only against recorded evidence in the log below, never against an argument. Refuted and supported name a scope: a hypothesis refuted for one provider or for one question stays open for the others, and the scope column is the load-bearing part.

Statuses: open (never tested), partial (tested in one scope only), supported, refuted, running.

The labels

Everything else in this register is conditional on these. They were assumed for months and tested for the first time on 2026-08-15.

#	Hypothesis	Status	What moves it
L1	A missing watermarked.unbound assertion means no watermark was embedded	refuted for both providers	Settled 2026-08-16. OpenAI: the official verifier called 65 of 94 non-asserted rows detected, a contamination rate of 69%, with 9 of 9 interleaved health positives detected so the verifier was demonstrably answering throughout. Google: its own API emits watermarked images, per documentation that states so outright, carrying no assertion. Every corpus-labelled AUC in this document rests on a negative class that is roughly two-thirds positive
L2	A watermark assertion means a pixel watermark is actually present	partial	Ten source controls returned detected across both providers. Ten of twelve thousand is a smoke test, not a rate. As of 2026-08-16 the verified-positive count is past thirty: Google web-verifier source arms across three accounts, the OpenAI provenance-API batches v1 and v2 (every asserted source detected), and 9 of 9 interleaved health positives. Still no denominator, so still not a rate. Extending the same runner to a frozen sample of asserted rows would give one. Update 2026-08-19: the first health miss (21/22) resolved as a bad label, not a verifier flip; the row originals/2026-06-25/198ae69e... is classified openai/asserted yet carries claim_generator: Google C2PA Core Generator Library and Google-side geometry, and the verifier's not_detected is the correct answer for it. The health design caught a corpus labeling error, which is the first direct evidence that assertion rows also carry at least one mislabeled member
L3	Date and model version do not carry the label	supported	The matched sets fix size, generator and codec but not date, and 57 of 70 Google controls fall in one month. Tested directly on 2026-08-16: a date-only classifier scores AUC 0.524/0.491, so date carries nothing and the covariate re-run is no longer load-bearing

Where the signal lives

#	Hypothesis	Status	What moves it
S1	The periodic comb carries the watermark	refuted	Settled 2026-08-16. The structure every periodic expert reads is destroyed by a seven-pixel crop while the published mark retains 99.97% TPR under aggressive crop and resize, so it is a generation-pipeline lattice anchored at the image origin, not the watermark
S2	The signal is the broadband content between the comb lines	partial	First direct oracle arm 2026-08-15: the equal-RMS broadband randomization (54.27 dB, local score unchanged by construction) remained detected by the Google verifier, as did every other arm of that account-isolated group. One image group at one perturbation class is a bound, not a sweep, but the between-lines content was randomized without moving the verdict, which is evidence against this scope. The local readout is blind between the comb lines by construction, so no local rerun can substitute; a wider amplitude ladder needs new provider quota
S3	The signal is the shape of the normalized mid-band spectrum	supported for OpenAI, open for Google	OpenAI confirmed under nested family and lambda selection on 2026-08-15: 0.510 shipped to 0.605, CI [+0.045, +0.147], above the permutation null. The Google branch was tested the same way on 2026-08-16 and did not improve: 0.678 against 0.689 shipped, CI [-0.071, +0.046]; at 63 controls a 0.02-0.03 gain is unprovable there in principle, so Google stays control-limited
S4	The carrier is content-dependent, so no fixed key-derived pattern exists	refuted for Google, supported for OpenAI	Settled 2026-08-15 by cross-image tile correlation. Google carries one fixed phase-coherent pattern, present in the controls at about half the amplitude; OpenAI carries none at chance level. What remains open is whether the Google controls are unwatermarked at all, which is L1
S5	The signal needs a learned nonlinear representation, not hand statistics	refuted at the tested capacity	First slice 2026-08-17: a 4-layer conv on opponent-residual patches, translation-invariant by construction (unaligned random patch offsets), leave-one-era-out. Against photographs it reaches 0.6-0.94, against foreign generators 0.44-0.56 with reversals to 0.11-0.15 on fresh eras, at a permutation null of 0.54-0.59: it learns "AI render versus photograph", nothing vendor-specific. A larger model remains free to try but inherits the same leave-one-era-out and leave-one-vendor acceptance gate
S6	Some hand-crafted statistic of the patch bank separates in every era, vendor and codec cell	refuted	Settled 2026-08-17 by the full-grid audit: of 124 features, zero clear minimum AUC 0.55 over 27 cells (best worst-cell 0.481), and a corpus-fitted ridge transfers to fresh eras at 0.388 native, 0.178 after JPEG-95

Robustness

A detector that fails these is not shippable whatever its AUC.

#	Hypothesis	Status	What moves it
R0	A candidate expert survives a crop that is not tile-aligned	refuted for every expert built so far, including the shipped registered-v3 default	A seven-pixel offset takes the red-green agreement from 0.962 to -0.111 at every size on large-v1, and on 2026-08-16 a two-pixel diagonal crop killed every one of the 36 registered-v3 detections across 28 foreign-generator acceptances and 8 Google provenance positives. This is now a precondition for calibration, not a later robustness check

#	Hypothesis	Status	What moves it
R1	The patch-local feature survives JPEG, resize and crop	partial, and its numbers await clean labels	Re-score the matched set under each transform. The earlier patch-bag model kept one positive in 486 after JPEG-95, so this is a gate and not an improvement
R2	The OpenAI result is not an artifact of PNG encoding or one geometry	supported, on contaminated labels	Repeat the evaluation inside a single exact size and a single month

Scope and universality

#	Hypothesis	Status	What moves it
U1	One feature space serves both providers	refuted at the current features	Transfer scores 0.535 and 0.469. A shared representation, or an honest per-provider router with abstention
U2	A watermark detector does not fire on non-watermarked AI images from other generators	refuted	Measured 2026-08-15 at feature level (the Google shipped feature accepts 34% of Adobe Firefly output) and 2026-08-16 through the production entry point: 29/223 signed non-Google generator images accepted, 0.130, Firefly 0.241. Photographic controls cannot bound this rate; see the two empirical entries below
U3	The detector works on the distribution it will actually meet	open	Both classes are drawn from files that still carry C2PA. A re-saved image carries none, and that is the case a user brings

Method validation

These do not test SynthID. They test whether our method can find a watermark it is known to contain, which is what separates "the features are weak" from "the data is too small".

#	Hypothesis	Status	What moves it
M1	The pipeline recovers a watermark we embedded ourselves	supported for a fixed-structure mark at SynthID's amplitude	Open watermarkers, VideoSeal, Watermark Anything, Stable Signature, give unlimited matched pairs. A pipeline that misses those has a method problem, not a data problem. The 2026-08-15 dwtDct ladder adds the boundary: full strength AUC 0.990, quarter strength (RMS 0.563, our SynthID measurement range) 0.936, so the pipeline is sound at the amplitudes we work at. The M1b/M1c ablation the same day bounds every hand-statistic branch: at equal RMS a fixed mark scores 0.670, a keyed mark 0.669, and a keyed perceptually masked mark 0.546, and demasking does not recover it (0.538). Masking, not keying or content dependence, is what defeats hand features
M2	Matched pairs can be minted rather than mined	refuted	Imagen, whose addWatermark the plan was built on, is absent from the model garden and every id 404s. The gemini-*-image models that replaced it reject the parameter, and the documentation states that all generated images carry a SynthID watermark. No current Google path can emit an unwatermarked image. Re-verified independently on 2026-08-16: imagen-4.0-generate-001 and imagen-3.0-generate-002 return 404 in us-central1 and us-east1 on both GCP projects, gemini-2.5-flash-image rejects addWatermark at request parsing (Cannot find field), and the current Gemini API documentation states for both the Gemini-image and Imagen paths that all generated images include a SynthID watermark, with no disable parameter documented. Settled on live API and documentation. Extended 2026-08-17 by the attack routes on the OpenAI side: pixel attacks to PSNR 20.1 (JPEG q40, 0.5 resize, sigma-8 noise plus JPEG q85, 0.35 resize plus JPEG q50) and a foreign-VAE decoder-substitution round-trip to PSNR 22.3 all left the official verifier answering detected, so attacking a verified positive cannot mint a verified negative on either provider by any mechanism tried
M3	folded_template_score is correct	supported	Settled 2026-08-15: an independent reimplementation of the fold scored over the same sample agrees with the shipped implementation to 3.7e-4, the residual traced to float32 in the cv2 blur. Removal and detection results still share the one implementation, but it now matches an independent one

The shipped decision

#	Hypothesis	Status	What moves it
D1	Some usable fraction of images can be decided at a precision-first threshold	refuted at one percent control acceptance	Asked and measured 2026-08-16: at one percent control acceptance the decidable positive fraction is 0-4.5%. A precision-first threshold that decides almost nothing is not a usable product surface

Empirical log

2026-08-09: D1 confound pilot and codec challenge

The manifest-driven D1 harness was run on the frozen Google pilot containing five oracle-positive images and 330 deduplicated exact-geometry external photographs. The manifest passed byte, decoded-pixel, lineage, and split verification. It is not evidence-ready because the locked test contains no same-provider hard negative.

The container-only baseline separated the labels perfectly because the pilot still exposes format and export-geometry differences. That result is a measured confound, not watermark evidence. The canonical 8x8 decoded-content baseline was much weaker: locked-test AUC was 0.620 and temporal AUC was 0.533, and its validation-frozen threshold detected neither held-out positive. The existing positive-only RGB plus HSV S/V ensemble still detected all five positives and emitted no positive verdict on the 330 frozen external images.

A codec challenge re-encoded every positive without changing geometry. The ensemble remained positive on five of five JPEG-95 outputs and five of five WebP-95 outputs; JPEG-90 retained three of five. This rejects a bare PNG versus JPEG container explanation, but does not exclude a generator or source-pipeline correlate. With only five positives, the one-sided 95% lower bound on TPR is 54.9%. With zero positives among 330 external images, the one-sided 95% upper bound on FPR is 0.904%, still nine times the 0.1% detector target. Four positives also participated in model fitting, leaving only one independent temporal positive. The next valid detector claim still requires new oracle-positive images and ordinary same-provider oracle-negative controls.

A post-freeze source-provenance challenge then added three exact-geometry Google originals that had not influenced fitting or threshold selection. Each carried the same signed Google LLC C2PA issuer, trained-algorithmic-media source type, and explicit SynthID-present assertion as the detected temporal control. The ensemble abstained on all three: two lacked active carrier support in both branches, and the third passed RGB evidence but missed the HSV S/V evidence and support gates. These are provider-signed embedding assertions rather than matching-oracle pixel labels, but zero positives in three new same-geometry images falsifies the current ensemble as a general Google SynthID detector. The measured phase family is retained only as an epoch- or surface-specific correlate pending a broader oracle-labeled corpus.

A follow-up cross-epoch check found that the correlate is not entirely confined to the original five images. RGB and HSV models fitted only on the three later images ranked each of the five earlier images above all 329 external negatives in both color spaces (AUC 1.0), but thresholds derived from the later fitting scores transferred poorly: the strict RGB-plus-HSV decision retained only one of five earlier positives at zero false positives on the 279-image holdout. An eight-fold leave-one-positive-out refit across both source groups then detected six of eight excluded positives with the same RGB-plus-HSV conjunction and each branch frozen just above its 50-image calibration maximum. RGB alone put all eight excluded positives above that maximum, while HSV missed the geometric and low-texture images; the per-fold RGB operating points also produced between one and eight false positives on 279 held-out negatives. This is evidence for a transferable but content- and epoch-sensitive Google pixel correlate, not a shippable detector. The experiment still lacks ordinary same-provider oracle negatives, and its external-negative set is too small for the 0.1% false-positive target.

2026-08-09: fixed spectral-template baseline rejected

An exploratory Google template was reconstructed from four public, purported clean/marked pairs. Their provenance and oracle status could not be established, so they were used for discovery only. A scalar phase-consensus score and its threshold were selected on those pairs, four older Google-oracle positives, and 30 external negatives.

The frozen threshold then produced six false positives on a new 100-image external holdout. It also detected only one of four newly generated Gemini images collected after threshold selection. Duplicate-image removal cannot reduce the false-positive rate enough to approach the 0.1% detector gate, and the temporal result is far below the required sensitivity. The baseline is therefore rejected, not recalibrated on the holdout.

This result rules out the fixed phase template as a detector or removal loss. It does not rule out a content-adaptive or model-epoch-specific signal. The next detector must learn from independently labeled provider data, must retain the failed holdout unchanged, and must demonstrate discrimination from export format and generator identity.

2026-08-09: cross-color low-texture consensus rejected

A polarity-invariant consensus template was trained on black, white, and red Gemini low-texture probes and frozen before evaluation. Its median score fell from 0.574 on the training groups to 0.354 on held-out probes of the same colors and 0.011 on unseen blue, green, and gray probes. External negatives had a median score of 0.0047. The separation therefore depended on the training colors and did not generalize across the intended low-texture stratum.

This branch is rejected as a current Google detector. It remains useful as a negative control demonstrating why high training coherence is not evidence of a shared carrier.

2026-08-09: external V3 codebook detector pilot

The numeric format-v2 V3 artifact from reverse-SynthID was loaded with pickle disabled and evaluated independently; no third-party code was imported or executed. A 256-bin phase score with a discovery-frozen threshold of 0.5 detected five of five Google-oracle-positive images and produced zero false positives on 194 external images collected before and after threshold freeze. The pilot includes four older positives and one newly generated temporal positive. It is promising discovery evidence, not a released detector:

• zero errors on 194 negatives cannot support a 0.1% FPR claim;
• the set lacks enough same-provider, oracle-negative hard controls;
• the phase profile is a third-party artifact whose positive provenance is not independently reproducible from our current corpus;
• a secondary active_weight_fraction shortcut failed on an exact-resolution holdout, firing on 17 of 50 new negatives, and is rejected as a native-resolution/content confound.

The next detector milestone is to reproduce the phase carrier from our own deduplicated positives, expand the negative set to at least the rule-of-three minimum, and add same-provider hard negatives before fitting or publishing an operating point.

2026-08-13: external V4 cross-color codebook audit

The later aloshdenny/reverse-SynthID V4 artifact was audited independently at commit b110836. The audit read only numeric NPY members from its NPZ ZIP with pickle disabled and reproduced the published native-resolution phase score; no third-party Python or pickle payload was executed. Its bundled sample_watermarked.png crossed the documented phase_match > 0.52 threshold at 0.535, confirming that the reproduction follows the intended path.

The operating point did not transfer. Taking the better of the two published model profiles detected 141 of 355 Google positives and 191 of 499 controls. On a later frozen 1,000-image Open Images reserve it accepted 386 controls. The phase score had AUC 0.517 on the paired corpus. Reproducing the July 2026 GUI's 0.3%-aspect-ratio gate reduced coverage without adding discrimination: it supported 121 positives and detected 46, while supporting 343 controls and detecting 132; it also detected 284 of 685 supported fresh controls.

The failure follows directly from the metric. Per-bin phase similarity is 1 - abs(wrapped_difference) / pi, whose expectation under independent uniform phase is 0.5, only 0.02 below the threshold. The detector selects bins by cross-color phase coherence alone and does not require image-side carrier amplitude or a content-baseline excess. Its selected bins span nearly the full spectrum, with a typical spatial period around 2.5 pixels. Cross-color coherence is useful for discovering candidates from genuinely low-texture references, but this single-image score is not a calibrated detector. Two prespecified rerankings did not rescue it. Restricting carrier periods to 4-128 pixels and ranking by either coherence times reference magnitude or the stored carrier weight times reference magnitude produced best-of-model AUCs of 0.521 and 0.524. They were evaluated as continuous scores without selecting a new test-set threshold and are also rejected.

The repository's older V3 phases also failed this external challenge. Their best dark/white phase score had AUC 0.473; the documented 0.78 phase threshold accepted 5 of 355 positives, 5 of 499 controls, and 6 of the later 1,000 fresh controls. The repository's current bundled sample and four validation sources all remained below that phase threshold. The README's accuracy claim is therefore not accepted as evidence for the current corpora.

The linked Hugging Face dataset cannot repair OpenAI calibration as published. Its DALL-E 3 black and white buckets are not solid-color probes: per-image mean luminance ranged from 11.66 to 143.74 and 124.94 to 235.66 respectively. Across ten 1024-square images per bucket, only 39 and 32 RGB frequency bins had phase coherence above 0.9, versus 20,443 and 100,252 for the corresponding Gemini 3.1 buckets. Only four DALL-E bins cleared 0.9 in both colors and all four were trivial DC or self-conjugate-axis coordinates; Gemini had 11,478 nontrivial joint bins and 3,341 also had cross-color absolute-cosine agreement above 0.9. Those DALL-E folders measure heterogeneous content phase, not an OpenAI solid-reference codebook. The official SynthID-Image paper also says nearly uniform corner cases may be excluded from watermarking, so generating flatter references does not guarantee a stronger watermark. A useful next external experiment requires newly collected low-content OpenAI outputs whose pixel watermark is separately confirmed, plus same-provider oracle-negative controls, before freezing any phase bins.

2026-08-09: independently fitted phase-carrier pilot

An exact-geometry phase model was then fitted from four of our own Google-oracle-positive images. The first unconstrained search was a clear multiple-comparisons failure: training scores were 0.983-0.991, while the newly confirmed fern positive scored 0.061 and an exact-resolution negative reached 0.105. Leave-one-out discovery detected only two of four older positives. That unrestricted model is rejected.

A second model used only the 256 frequency coordinates proposed by the numeric V3 artifact, while relearning phase, weight, and expected magnitude entirely from our four positives. This restricted model scored the fern validation positive at 0.740 and the four fitting positives at 0.936-0.963. A threshold of 0.43 was frozen after 50 validation negatives whose maximum was 0.305.

The phase-only epoch failed its first new holdout with three false positives in 121 exact-size images. All three had active carrier weight between 0.07 and 0.14, so the failed epoch was closed and a new rule was frozen: phase score at least 0.43 and active weight at least 0.50. That second epoch produced zero false positives on 279 newly downloaded exact-size images.

A separate 3,000-image COCO check also produced zero joint-rule positives, but it is not counted toward the supported-stratum FPR bound. Every COCO image had been upscaled to the model geometry and every active weight was at most 0.201. The rule therefore abstained because the inputs were outside its measured carrier-support distribution. Treat active weight as an abstain gate, not as negative SynthID evidence.

This is the first positive local-detector pilot derived from our labels, but it does not clear the detector gate. It still depends on third-party candidate coordinates, has only five confirmed positives including the fitting images, has only 279 fresh negatives inside the current support stratum, and has no same-provider hard negatives. The next epoch requires a larger independently labeled positive set, at least 3,000 native-support negatives, same-provider oracle negatives, and a new temporal positive that has not influenced feature or threshold selection.

2026-08-09: color-space spectral comparison

The restricted phase-carrier experiment was repeated in RGB, full-range YCbCr, YCoCg, an orthogonal opponent basis, CIE Lab, and HSV. Every branch started from the same 102 spatial frequencies: the top 256 external codebook coordinates contained 102 unique (row, column) pairs, which were expanded over all three components. Each branch then independently selected 256 of the 306 component-frequency candidates and relearned phase, expected magnitude, and weight from the same four oracle-positive fitting images. The score was the phase score multiplied by active carrier weight, so weak spectral support reduced rather than merely qualified the evidence.

Thresholds were frozen from the fern validation positive and 50 exact-size validation negatives before scoring the 279-image comparison set. The comparison set was locked for this color-space branch, although it had already served as the second epoch's RGB negative set and is therefore not a globally virgin corpus.

Space	Validation gap	Comparison negative max	Fern minus negative max	False positives
RGB	0.509	0.097	0.459	0/279
YCbCr	0.330	0.193	0.312	0/279
YCoCg	0.437	0.163	0.390	0/279
Opponent	0.351	0.167	0.339	0/279
Lab	0.408	0.126	0.382	0/279
HSV	0.530	0.084	0.495	0/279

HSV had the best observed worst-negative margin, narrowly ahead of RGB, but did not generally shift paired negatives below RGB. After normalizing each space by its fern score, the median paired HSV-minus-RGB difference was 0.0015 and a two-sided sign test gave p=0.632 (144 higher, 135 lower). The result is therefore a tail observation, not evidence that HSV dominates RGB.

Channel decomposition localized the useful effect. HSV hue contributed only 0.012 of fern evidence and its fern-minus-holdout-maximum channel gap was -0.024; saturation and value contributed 0.333 and 0.233, with positive gaps of 0.273 and 0.194. In YCbCr, YCoCg, opponent, and Lab, the luminance-like channel supplied the largest and most stable separation, while chroma-channel gaps were much smaller. This supports a carrier expressed across intensity and contrast structure, with useful nonlinear saturation/value projections, not a hue-specific watermark.

This experiment does not raise the detector's evidence grade. It still has only one held-out confirmed positive, reuses the same third-party spatial frequency prior, lacks same-provider hard negatives, and evaluates only one native geometry. A future HSV S+V ablation must freeze its own operating point and use a new negative epoch and new oracle-positive images; dropping hue now would be a post-holdout model choice.

2026-08-09: pixel-only R1 pilot did not transfer

Three non-generative candidate families were frozen locally before Gemini verification:

1. local visible-region inpainting changed 0.087% of pixels at 47.64 dB PSNR and 0.99963 SSIM;
2. sparse complex-spectrum subtraction used 256, 4,096, or 16,384 numeric V3 bins while preserving native 2816x1536 geometry;
3. deterministic spatial fragmentation combined smooth local warps, one global resampling operation, mild color changes, and codec round-trips.

Local inpainting left the 256-bin phase score effectively unchanged. Healthy Gemini sessions continued to detect both the target and its spatially shifted sham, refuting an earlier one-off not detected response.

The widest spectral candidate reduced the frozen 256-bin phase score from 0.639 to 0.124 while retaining 51.94 dB PSNR and 0.99854 SSIM. A pixel-identical, metadata-normalized source control was detected in two healthy oracle sessions. The 16,384-bin candidate remained detected in one of them and the second hit quota before returning a verdict. The 4,096-bin candidates also remained detected. Therefore the external phase carrier is correlated with the Google signal but is not a sufficient removal loss.

One additional profile returned not detected for the 16,384-bin candidate, then also returned not detected for the untouched source. That entire session is invalidated rather than counted as removal success. This failure adds a new oracle-health rule: every account/session batch starts with the same pixel-identical positive control; candidate verdicts are discarded when that control is not positive. Quota and tool-unavailable responses remain indeterminate.

The frozen fragmentation batch is not yet oracle-resolved because healthy profiles exhausted their rolling quota. Its balanced and strong variants also fall below the release fidelity gate locally, so even a future negative oracle result would be a mechanism-discovery result, not a releasable remover.

A follow-up bounded-warp batch replaced RMS-normalized displacement with strict per-axis maxima of 1.0, 1.8, and 2.8 pixels. The 1.0-pixel variant retained 37.79 dB PSNR and 0.9784 SSIM but left the local phase score essentially unchanged at 0.643. The 2.8-pixel variant reduced the score only to 0.503 while falling to 28.21 dB and 0.8570 SSIM. A multi-transform bounded variant reached 0.159 locally only at 21.97 dB and 0.6749 SSIM. Spatial fragmentation therefore does not approach the release fidelity gate in this implementation and remains a preregistered mechanism probe for the next healthy oracle window, not a product candidate.

2026-08-09: positive-only RGB plus S/V detector epoch

The color-space result was converted into a deliberately one-sided research detector. It requires exact 2816x1536 geometry, RGB evidence of at least 0.302, HSV saturation-plus-value evidence of at least 0.302, and active carrier support of at least 0.50 in both branches. It emits positive only when every condition passes and emits abstain otherwise; it never claims that SynthID is absent.

Leave-one-positive-out fitting passed all four older Google positives. Their minimum held-out RGB evidence was 0.527, minimum held-out S+V evidence was 0.452, and minimum active support was 0.600. The final frozen models detected all five available oracle-positive controls. A new Picsum epoch downloaded IDs 600-999; after exact-geometry validation and content-hash deduplication against all earlier epochs, 330 images remained. None produced a positive verdict. However, 328 of 330 abstained for insufficient carrier support and only two reached both support gates before falling below the evidence thresholds. This validates a positive-only signal catcher inside one geometry, not a two-sided detector or a general FPR claim.

2026-08-09: ensemble phase projection is not a sufficient removal loss

An analytical pixel-only attack removed the positive complex projection onto the learned RGB phases and HSV S/V phases. One iteration at strength 0.5 cleared the local ensemble at 56.03 dB PSNR, 0.99909 SSIM, and a maximum pixel change of two levels. A norm-matched random-noise control remained locally positive.

The provider result rejected this loss as sufficient. In a Gemini session where the untouched fern control returned SynthID Status: Detected, all three preregistered projection candidates at strengths 0.5, 0.75, and 1.0 also returned detected. The strongest candidate had negative local RGB and S+V evidence at 51.97 dB PSNR. The ensemble therefore tracks a real correlate but does not cover all evidence used by Google's decoder.

2026-08-09: 16x32 periodic-tile carrier hypothesis

The failed global projection exposed a more specific spatial structure. Of the top 256 external V3 bins at 1536x2816, 208 (81.25%) lie exactly on an FFT lattice spaced by 96 rows and 88 columns. This corresponds to a spatial cell of 1536 / 96 = 16 rows by 2816 / 88 = 32 columns. Instead of subtracting a small list of harmonics, a new pixel-only probe high-pass filters the image, folds the residual modulo 16x32 over 8,448 repeats, and subtracts the complete zero-mean folded template.

On the fern positive with its visible sparkle already removed, the mildest locally clearing candidate used Gaussian sigma 0.6 and strength 1.5. The phase-aligned template cleared the ensemble at 57.03 dB PSNR and 0.99911 SSIM; the same template shifted by one row and one column remained positive at 57.02 dB and 0.99912 SSIM. Two additional aligned/shifted pairs reproduced the direction at 55.96 and 54.19 dB.

The per-image estimator was then run on the four older positives with a frozen strength ladder. The minimum locally clearing strength was 2.0 for three images and 1.5 for one. All four phase-aligned outputs abstained, while all four same-strength one-pixel-shifted controls remained positive. Selected outputs retained at least 55.67 dB PSNR and 0.99874 SSIM. This is the strongest local causal evidence so far because effect direction follows exact tile phase at matched distortion across five contents.

It is not a removal result. In a healthy Google verifier session, the fern source control was detected and the mild aligned sigma-0.6, strength-1.5 candidate was also detected. The stronger local variants therefore remain unresolved rather than inheriting a negative label from the local ensemble. Cross-resolution tile discovery and same-provider hard negatives also remain open.

2026-08-09: provider-specific phase losses failed both remote verifiers

The Google result above and a separate OpenAI experiment closed the first analytical-loss epoch. The OpenAI branch fitted an 8-pixel phase pattern from OpenAI low-texture outputs and produced a 46.70 dB candidate whose local phase score changed sign. A single frozen submission to OpenAI Verify still returned SynthID detected; C2PA was absent because the candidate had been re-encoded. This confirms that metadata removal and a local phase-score reversal do not establish pixel-watermark removal.

The experiments are not cross-provider replications. Google images were sent only to Google Verify AI and the OpenAI candidate only to OpenAI Verify. No carrier, threshold, response, or verdict was transferred between tracks. OpenAI's published verifier guidance prohibits repeated queries for watermark reverse engineering or evasion, so that one failed query did not start an adaptive loop. Further OpenAI candidates remain local hypotheses unless an authorized research oracle becomes available.

2026-08-09: JPEG-robust periodic EOT surrogate pilot

Small residual CNNs were trained separately for each provider as differentiable attack surfaces. They are provider classifiers, not SynthID detectors: their corpora lack current same-provider oracle-negative controls, and several external negatives approach the positive range. Two same-architecture models were used for optimization and a third model with a different first kernel and channel schedule was kept outside the attack.

For Google, all three models scored the temporal fern positive at a 75th-patch percentile between 0.938 and 0.962. A patch-local L-infinity-8 attack initially cleared the models and survived JPEG, but visual inspection exposed rectangular grid boundaries; that candidate was rejected despite favorable scalar metrics. A replacement optimized one 16x32 residual cell repeated over the full image, with differentiable JPEG-95 and JPEG-90 branches. At half strength it retained 36.80 dB PSNR and 0.9241 SSIM against the icon-clean source. The three PNG scores fell to 0.024, 0.047, and 0.016; after an actual JPEG-90 round trip they were 0.182, 0.243, and 0.016. The independent RGB and HSV S+V phase ensemble also abstained, and remained below threshold after JPEG-90, JPEG-95, WebP-95, and a 0.98 resize round trip.

For OpenAI, the held-out complex market image scored 0.936, 0.929, and 0.813 on its three provider classifiers. An independently optimized 8x8 periodic EOT residual was then attenuated in low-texture regions. The selected local candidate retained 34.97 dB PSNR and 0.9356 SSIM. Its PNG scores were 0.030, 0.053, and 0.099; after actual JPEG-90 they were 0.042, 0.063, and 0.178. WebP-95 behaved similarly, while a 0.98 resize raised the independent model to 0.692 and remains a measured weakness. A second optimization that included a resize branch improved the two optimized models but transferred worse to the independent model, so it was rejected.

The periodic residuals were then tested for cross-image transfer without further optimization. The OpenAI 8x8 cell learned from the market image was applied with the same low-texture attenuation to 11 other OpenAI positives. Every image moved downward on all three models, both as PNG and after an actual JPEG-90 round trip. Median PNG scores changed from 0.971, 0.970, and 0.931 to 0.016, 0.033, and 0.036. Each model had 11 of 11 paired decreases, giving a two-sided exact sign-test p=0.00098; per-image PSNR ranged from 33.50 to 35.99 dB. Two complex scenes retained materially higher scores on the independent model, so the transfer is strong but not uniform.

The Google 16x32 EOT cell was likewise applied to the four older images after their own locally clearing phase-tile subtraction. All four remained phase ensemble abstentions, and every CNN score decreased in the lossless outputs. JPEG-90 preserved the direction but restored the two optimized models to roughly 0.58-0.78 on the hardest images. Fidelity was approximately 37.25 dB PSNR. With only four paired images this is descriptive transfer evidence, not a statistical or oracle gate.

These are positive local robustness results, not watermark-removal results. Neither candidate meets the release fidelity gate, neither CNN ensemble has cleared the detector-identifiability gate, and neither selected EOT candidate has a matching-provider negative oracle verdict. A three-file Google batch is hash-frozen outside the repository for the next healthy quota window: the positive source, the stronger phase-only ablation, and the combined periodic EOT candidate. Results will be accepted only in that fixed order without between-query adaptation. Two later attempts detected the positive source, then returned a usage-limit response for the phase-only ablation. In the latest attempt the verifier requested a retry after 17 hours. Both responses are indeterminate, and the combined candidate was not submitted out of order. The OpenAI candidate is not scheduled for the public verifier under the current usage restriction.

2026-08-09: OpenAI fidelity and architecture-transfer frontier

Further local experiments tested whether the OpenAI periodic result could move above the release fidelity gate without losing transport robustness. Chroma-only projection reached 47.43 dB PSNR, but an actual JPEG round trip removed the perturbation and restored the two classifier scores to approximately 0.9. Mixed luminance/chroma projections and stronger texture masks did not improve that frontier. A regularized Adam attack reached 41.21 dB on the source image, but JPEG-90 restored the optimized classifiers to 0.90 and 0.91, and the independent architecture scored 0.89. This rejected post-hoc color projection and simple regularization as fidelity remedies.

A second experiment put actual OpenCV JPEG and WebP results in the forward pass and used a straight-through gradient. At 40.08 dB, actual JPEG-90 still scored 0.89 and 0.91 on the optimized models. Removing the luminance penalty improved the first model to 0.76 but left the second at 0.88 and reduced fidelity to 38.78 dB. The exact-codec forward pass therefore corrected an invalid transport approximation, but did not produce a new Pareto candidate.

An architecture-diverse fourth classifier used fine RGB residuals, broader RGB residuals, and luminance/opponent-color residuals. The market image remained outside its training split. Its 75th-patch percentile ranged from 0.983 to 0.988 on four held-out OpenAI positives; the median across 120 held-out COCO negatives was 0.041 and the maximum was 0.521. This is useful attack-surface diversity, not a certified detector, because no current same-provider oracle-negative corpus is available.

Adding that model to periodic optimization did not improve architecture transfer. The preregistered local winner retained 34.86 dB and scored 0.057, 0.197, and 0.021 after JPEG-90 on the three optimized models, but the independent model scored 0.573. A per-image half-resolution perturbation produced a higher-fidelity 41.86 dB candidate; JPEG-90 scores on its optimized models were 0.587, 0.572, and 0.471, while the independent model remained at 0.875. Optimizing only the two original models produced the same failure on both independent architectures.

The measured local frontier is therefore unchanged. The 34.97 dB periodic candidate remains the strongest transport-robust and cross-architecture result; the candidates above 40 dB are architecture-specific and cannot support a removal claim. No additional public OpenAI verifier requests were made during these experiments.

2026-08-09: mechanism reset and paired open-method control

The first detector epochs optimized the wrong statistical problem. The technical reference now records the paper's paired training target, symmetric transformation channel, independent detection logit, and two-sided calibration. Applied to the initial pilots, those details reject three implicit assumptions: that a carrier phase can stand in for the detection score, that mean patch classification is an adequate image-level decoder, and that a sigmoid threshold calibrated on unrelated negatives is a deployable decision rule.

A bounded open-method control tested the architecture concern without claiming that another watermark reproduces SynthID. TrustMark P was used only because its public encoder creates causal clean/watermarked pairs and its public decoder can validate transformed labels. Ninety public COCO images were split by source image into 60 training, 15 calibration, and 15 test pairs. Each pair received the same identity, JPEG-90, 0.8 resize round trip, or 5% crop transformation. The official decoder detected 14/15, 15/15, 14/15, and 15/15 transformed marked test images respectively, with zero detections on the corresponding clean images.

The existing normalized residual patch recipe reached only 0.707 identity AUC and 0.668 aggregate AUC on the held-out pairs. A small full-field model that retained RGB amplitude plus fine and coarse residuals reached 0.947 identity AUC and 0.841 aggregate AUC. At a threshold above every calibration negative, their aggregate TPRs were 5% and 20% respectively. Both models failed to reproduce the official decoder's JPEG robustness. This small proxy is not a TrustMark benchmark and says nothing directly about SynthID accuracy. It does falsify the claim that the current patch-normalized architecture and training recipe are an adequate generic neural-watermark detector.

The resulting eight falsifiable hypotheses now define the canonical D3-D5 mechanism program. In particular, external generator corpora remain a final false-positive challenge and do not substitute for counterfactual negatives.

Until the counterfactual-label and full-field gates pass, the current residual CNN remains a useful vendor-triage stage for a future cascade, but it is not a SynthID detector and should not be optimized as the final decision surface.

2026-08-09: paired wavelet and spectral ablation

The paired spectral harness was extended with three-level undecimated db2 wavelets, complex Fourier phase coherence and power, cepstral peaks, and a cyclic clean/marked permutation control. The implementation streams wavelet and spectrum field accumulators rather than stacking transformed fields across pairs.

On 60 public TrustMark P training pairs at canonical size 256, the true residuals had mean inter-pair RGB NCC of 0.097-0.121. Cyclically mismatching each clean image with the next marked image reduced all three channels to about -0.017. The strongest coherent Fourier cluster occurred around vertical offsets 9-11 and was strongest in the blue channel. Wavelet repeatability was greatest in horizontal detail bands at levels 1 and 2. These are causal residual measurements for the open proxy, not evidence about SynthID's carrier.

The representations then faced a source-disjoint single-image test with 60 training, 15 calibration, and 15 test pairs under identity, JPEG-90, 0.8 resize, and 5% crop transformations. A fixed spectral-template score reached only 0.618 identity AUC and collapsed to 0.498 under crop. A 318-feature wavelet/spectral summary reached 0.653 aggregate AUC and zero TPR at the threshold above every calibration negative.

Full-field neural ablations at size 128 gave the following aggregate results:

Input	AUC	Paired wins	Test FPR	Test TPR
Spatial RGB plus residuals	0.723	57/60	0%	11.7%
SWT maps only	0.671	59/60	0%	0%
Complex FFT maps only	0.516	41/60	0%	0%
Spatial plus SWT, late fusion	0.665	59/60	5.0%	8.3%
Spatial plus SWT plus FFT, early fusion	0.639	57/60	1.7%	1.7%
Spatial plus SWT plus FFT, late fusion	0.698	54/60	0%	0%

The table uses one threshold above every calibration negative; test FPR is reported separately because a clean calibration result does not guarantee a clean locked test.

The SWT branch contains weak, unusually consistent paired evidence, but it did not improve the spatial model's source-disjoint discrimination in this small proxy. The FFT map branch was indistinguishable from chance. Therefore neither representation advances into the primary detector by default. SWT remains a late-fusion ablation for a larger paired corpus; complex spectral analysis remains a residual-discovery diagnostic unless a future held-out test reverses this result.

2026-08-09: closed-corpus architecture and score ablations

A temporally split closed three-class corpus then tested whether the open-proxy architecture result transferred to real external-source images. These labels establish source classes, not watermark presence; they can measure generalization and false positives but cannot substitute for counterfactual clean/marked labels.

The frozen normalized-patch baseline was repeated over three training seeds. Individual original-image AUCs ranged from 0.864 to 0.879. Averaging the three models increased AUC to 0.884, but a threshold above every validation negative still produced one false positive among 259 test negatives. At that operating point the two positive-source TPRs were only 2.45% and 8.97%. Calibrating near 1% validation FPR yielded 1.54% test FPR and TPRs of 6.68% and 18.16%. Accepted positives had 98.8-100% vendor accuracy, so vendor attribution is useful only after an independently reliable presence decision.

A naive full-field spatial model reached only 0.599 AUC. Adding a separately encoded stationary-wavelet branch and late fusion reduced AUC to 0.597. Both had effectively zero strict TPR. The result does not contradict the causal open proxy: it shows that global average pooling over source-labeled images does not isolate the transferable signal and that SWT does not repair the source shift.

Post-hoc image aggregation also failed to create an operating point. Averaging presence logits raised AUC from 0.884 to 0.890, but increased strict test FPR to 0.77% while reducing TPR. Median, lower-quartile, upper-quartile, and top-tail aggregation either preserved the overlap or suppressed almost every positive. A separately trained binary presence head plus vendor head reached 0.880 ensemble AUC, 0.77% strict test FPR, and TPRs of 1.11% and 7.26%. Decoupling the heads is architecturally faithful but insufficient without faithful labels.

These ablations close spectral feature engineering, naive full-field pooling, and score aggregation as the next detector step. The next valid training epoch requires same-content counterfactual pairs or independent watermark oracle labels, matched transformations, and an image-level detection loss. External generator corpora, including difficult non-target providers, remain hard negative and FPR-challenge sets only.

2026-08-10: low-content controls and registered phase carrier

A same-resolution low-content matrix compared independently generated solid outputs from two target model families against three per-image controls: exact mean fill, amplitude-matched Gaussian noise, and a phase-randomized residual with preserved Fourier magnitude. Raw stationary-wavelet summaries transferred between the two target families with AUCs of 0.982 and 1.000, and reached 0.973 when blue and green were held out by color. This was not watermark evidence. The frozen classifier accepted every one of 1,869 external negatives because it had learned the distinction between real generator texture and artificial controls. Removing absolute wavelet energy reduced external-negative acceptance only to 61.6%, with similar 58.1-67.3% acceptance across all three source classes. Both low-content wavelet branches are rejected as presence detectors until real non-target solid outputs provide matched negatives.

The numeric V3 audit loader was then extended to support both dense and sparse format-v2 profiles without pickle. Exact-profile evaluation exposed a sharp encoder-version boundary. The 1024x1024 profile accepted none of 231 target provider images and none of 26 exact-geometry negatives. The 1536x2816 profile accepted 30 of 55 target-provider images, including all four temporal-test images, while rejecting the one exact-geometry negative available in the closed corpus. The independently fitted phase model accepted 24 of those 55 and also accepted all four temporal-test images. This is positive evidence for a geometry- and epoch-specific carrier, not a universal SynthID decoder.

On the four temporal-test positives, the fixed V3 score survived JPEG-95 and a 75% downscale on all four images, survived JPEG-85 on two, and failed after a 5% center crop or a one-pixel translation on all four. Bounded analytical translation registration recovered all four shifted images and selected the known (-1, -1) offset. Searching up to 16 pixels produced no positives among 50 exact-resolution and 144 canonicalized frozen negatives. The shared registration implementation now serves both the numeric V3 probe and the independently fitted phase model.

A discovery-only scale-and-translation view search recovered all four 5% cropped temporal images with the independently fitted model after lowering the active-support gate from 0.50 to 0.40. It produced zero positives on the 194 frozen negatives and on the same preregistered 3,000-image COCO challenge used by the identity scorer. The latter result has a zero-error one-sided 95% bound of 0.0998% only for that abstention challenge: every COCO image remained outside carrier support, with a maximum active fraction of 0.201. The scale rule is not frozen because its support threshold was selected after inspecting the crop examples. It requires a new temporal positive holdout before it can join the detector rule.

The current actionable research candidate remains a positive-only, provider-specific expert for the supported 1536x2816 carrier epoch. Identity and bounded translation views use the frozen phase and support thresholds; unsupported geometry, insufficient carrier magnitude, and ambiguous phase return abstain. Vendor attribution may select the expert that supplied accepted evidence, but it must not turn an abstention into a provider label. The next calibration gate still requires at least 3,000 native-support negatives, same-provider oracle negatives, matched non-target solid outputs, and a new temporal positive that influenced neither profile nor threshold.

2026-08-10: 2048 periodic-tile detector

The phase hypothesis transferred to a second native geometry when the model and operating point were kept separate. A 256-peak 2048x2048 model learned from 111 train positives used the already frozen phase threshold of 0.43, active support threshold of 0.40, and a translation search of plus or minus four pixels. It accepted 31 of 35 validation positives and 27 of 30 locked-test positives while accepting none of 49 validation or 38 test negatives. Translation registration recovered the shifted positives without changing those identity counts. A threshold recalibrated from only 49 validation negatives reached 30 of 30 test positives but accepted one test negative, so it was rejected in favor of the transferred rule.

The wider native-geometry challenge exposed the remaining uncertainty. The frozen 2048 rule accepted two of 182 earlier external-provider images, for two accepted source negatives among all 269 native negatives. Both cases passed at zero translation with high phase and support, and both also passed an independently learned HSV phase branch. They are operational false positives under source labels, but source provenance does not establish watermark absence. They may instead expose a shared encoder or upstream backend. Without an independent watermark oracle they cannot be relabeled either way. The same experiment rejected the 1024x1024 and 768x1376 experts: they accepted 9 of 26 and 4 of 9 native source negatives, respectively.

The 2048 carrier has a concrete periodic mechanism. Its 256 peaks reduced to 108 unique spatial frequencies. Translating the frequency coordinates by 128 rows preserved 56 coordinates, while the maximum overlap in each of 1,000 uniform random controls was two. The permutation estimate was 0.001, and the 128-bin spacing implies a 16x16 spatial tile. A separate detector therefore folded a high-pass residual modulo 16x16, averaged 16,384 repetitions, and correlated the normalized tile against a train-positive template. After float64 serialization and validation-only threshold calibration, the fixed tile accepted 34 of 35 validation and 29 of 30 test positives, none of the 49 calibration or 38 held-out test negatives, and the same two of 182 earlier source negatives. This second representation supports a real periodic carrier rather than an arbitrary set of FFT peaks, but it does not resolve the two labels.

The fixed tile also accepted none of a preregistered 3,000-image general-image challenge after deterministic 2048x2048 canonicalization. The maximum normalized correlation was 0.106 against a threshold of 0.174. The zero-error one-sided 95% upper bound is 0.0998% for that challenge. Unlike the earlier sparse-phase result, the tile score evaluates every image without a weak carrier-support abstention. The challenge is still not native provider data and cannot replace the required oracle-negative calibration.

A later exact-1024 content audit found that perceptual siblings can cross an image-level fit split even when file and decoded-pixel hashes differ. The local 2048 reports retain aggregate counts and model hashes but not the 111 fitting paths, so the old train/validation/test split cannot now be audited for that specific leakage mode. This does not invalidate the independently collected general-image challenges or justify changing the frozen runtime model, but the reported positive rates remain conditional on the historical split. Any next model calibration must retain content-group membership and prove that no group crosses train, validation, test, or temporal partitions.

Symmetric attack evaluation established the robustness boundary. The fixed tile accepted 29 of 30 original test positives, all 30 after a 75% downscale round trip, 21 after JPEG-95, three after JPEG-85, and none after a 5% crop, with no accepted held-out negatives under the identity threshold. A validation-calibrated JPEG-95 tile threshold recovered 27 of 30 but accepted one of 38 test negatives. Requiring both codec-conditioned tile and phase scores reduced JPEG-95 to 16 of 30 with no held-out-negative acceptance, but still accepted one of the 182 earlier source negatives. The corresponding JPEG-85 consensus accepted 5 of 30 positives and none of all 269 native source negatives. Scale-and-translation phase search recovered 15 of 35 validation and 14 of 30 test crops with no held-out-negative acceptance, but remains discovery-only because the test transformation had already influenced the branch. Low-frequency peak subsets and transform-augmented phase training improved JPEG sensitivity only by raising validation false positives to 2-10%, so both were rejected.

The reproducible implementation is scripts/synthid_periodic_tile_probe.py. It stores the normalized template in float64 and calibrates only after loading the serialized artifact; an earlier float32 experiment moved a boundary score by approximately 2.5e-10 and demonstrated why calibration-before-serialization is invalid. The resulting research detector is positive-only and limited to a confirmed 2048x2048 carrier epoch. An accepted expert may suggest the encoder family, but the two cross-source carrier matches prohibit a stronger vendor claim until an oracle distinguishes direct provider output from shared-backend output.

A pixel-space ablation then tested whether the frozen tile merely predicted the local scores or controlled them. At twice the train-median tile norm, aligned subtraction changed the fixed-tile decision from 29 of 30 accepted originals to zero and the independently fitted sparse-phase decision from 27 of 30 to zero. The median fidelity was 53.74 dB PSNR and 0.99681 SSIM. Cyclically shifting the same tile by one row and column left one phase acceptance, while a seeded zero-mean random tile orthogonal to the learned template left 13; their median PSNRs were within 0.13 dB of the aligned edit. For both local representations, the aligned edit reduced the score more than either control on all 30 paired images, with a two-sided sign-test p-value of 1.86e-9 for each comparison. The same aligned edit reversed both local decisions on each of the two disputed cross-source matches, while the orthogonal control left both phase decisions accepted. That two-item result is descriptive but makes an accidental threshold crossing less likely; it still cannot distinguish a shared carrier from direct provider output.

This is the strongest local causal evidence for the 16x16 mechanism, but the strength was selected after inspecting this locked test and is therefore discovery-only. The shifted control also suppressed the phase representation substantially, so local score reversal cannot certify signal removal. A matching provider oracle must still compare untouched sources, aligned candidates, and norm-matched controls before any removal claim. The reproducible local harness is scripts/synthid_periodic_tile_ablation.py; its report records input and model hashes without writing derivative images.

2026-08-12: fast adaptive carrier projection

The fixed twice-median subtraction was intentionally conservative but spent substantially more pixel norm than many images needed. A follow-up tested a one-dimensional per-image solver. For source pixels x, frozen unit carrier t, and scalar amplitude a, the candidate is clip(round(x - a * tile(t))). A bounded scalar search selects the smallest a that reaches the local target. This is a direct pixel edit: it uses no generative decoder, does not resize the image, and searches one scalar rather than optimizing a full-resolution tensor.

The first challenge reused the 30 historical 2048x2048 causal-ablation sources. Three were not simultaneously positive under the fixed-tile and independently fitted phase rules and were excluded before counting a reversal. All 27 conjunctive positives cleared both local rules after adaptive subtraction. Median PSNR was 59.83 dB and median SSIM was 0.99890; the minima were 55.19 dB and 0.99705. Repeating the same selected amplitude with a one-row, one-column cyclic shift cleared zero of 27. Sixteen original images remained conjunctively positive after JPEG-95. The aligned candidates cleared all 16 after the same round trip, while the shifted controls cleared one. This improves materially on the earlier fixed-strength median of 53.74 dB without weakening the local causal control.

A three-image pilot also compared subtracting the complete per-image folded high-pass residual and a regularized inverse-high-pass reconstruction. Both could clear the two local scores, but their selected candidates retained only about 54-57 dB PSNR, compared with 59-62 dB for the frozen carrier projection on the same images. Removing the whole folded residual therefore spends pixel norm on content-correlated periodic structure and was rejected in favor of the narrow carrier direction.

The geometry challenge then selected three first_only carrier positives from each of the 20 most frequent native Spaces geometries. The edit repeats the same 16x16 carrier past the image boundaries and crops the repetition to the decoded dimensions, so it does not require dimensions divisible by 16. A zero score target cleared 60 of 60 aligned candidates and 5 of 60 matched shifted controls. Median PSNR and SSIM were 60.37 dB and 0.99919; the minima were 53.56 dB and 0.99581. These samples come from an already inspected detector corpus, so they measure mechanism transfer across geometry rather than fresh detector or removal generalization.

The zero target left one WebP-95 and eight resize survivors among transformed sources that were still locally positive. A frozen follow-up used a target score of -0.25. It cleared all 60 native candidates at median 57.70 dB PSNR and 0.99863 SSIM, with minima of 52.38 dB and 0.99443. Conditional on the matching transformed source remaining above the fixed carrier threshold, the aligned candidate cleared 32 of 32 JPEG-95, 13 of 13 JPEG-90, 8 of 8 JPEG-85, 15 of 15 WebP-95, and 54 of 54 0.75x resize round trips. The corresponding one-pixel shifted controls cleared 7, 1, 1, 5, and 17. The increased shifted-control rate is the measured price of adding transform margin.

The frozen -0.25 recipe was finally run once on one deterministic first-carrier positive from every represented native Spaces geometry. All 647 images at 647 distinct decoded sizes reached the target, with no maximum-amplitude failure. Median fidelity was 57.67 dB PSNR and 0.99867 SSIM; the worst case was 51.78 dB and 0.99400. Of 323 source images whose carrier survived JPEG-95, 323 aligned candidates and 58 shifted controls cleared after the same round trip. Natively, 153 of 647 shifted controls also cleared at this deeper margin. The consistent aligned result supports geometry-independent control, while the increased shifted rate shows that the conservative edit perturbs a wider neighborhood of the periodic carrier. Because the all-geometry corpus had already influenced detector research, it is a coverage challenge rather than a fresh holdout.

This is a fast local carrier suppressor, not a verified SynthID remover. The earlier Google oracle remained positive after analytical edits that already looked favorable to local models, and the adaptive scalar recipe has not been submitted to a matching-provider oracle. Provider terms and the project oracle policy also prohibit turning verifier responses into an adaptive search loop. The next valid test is one preregistered source/re-encode/aligned/shifted batch at a healthy authorized oracle, with the amplitude rule and -0.25 target frozen before submission. Until that batch returns a matching-provider negative result, no production command exposes this method as removal.

The reproducible implementation is scripts/synthid_adaptive_carrier_suppress.py. It writes only lossless PNG, refuses overwrite and locally negative input, and records source/output hashes, the selected scalar amplitude, local scores, fidelity metrics, and runtime in a sidecar JSON report. Unit tests cover non-divisible geometry, target attainment, an unreachable target, and the negative-input guard. A real CLI smoke test on a Spaces source reproduced its reported score exactly through the runtime detector after saving. It remains a research script rather than a routed package command.

2026-08-10: confirmatory oracle-batch readiness

The corpus audit found no new 2048x2048 positive collected after the tile rule and subtraction strength were frozen. Existing test images cannot be renamed as a temporal holdout, so the confirmatory oracle batch remains blocked on new provider output at the supported native geometry.

The batch mechanics are ready in scripts/synthid_oracle_batch.py. For each new source, it preregisters five requests in a fixed order: untouched source, lossless re-encode, aligned subtraction, cyclic-shifted subtraction, and an orthogonal norm-matched tile. It refuses to write media inside the repository, records artifact and decoded-pixel hashes, pins both model artifacts and the matching provider, and detects any post-registration byte change. Oracle results live in a separate manifest-bound file so the preregistration remains immutable; the evaluator requires complete ordered results, verbatim responses, timezone-aware timestamps, and separate SynthID and C2PA verdicts.

A causal success requires the source, re-encode, shifted, and orthogonal views all to remain detected while only the aligned view becomes not_detected. Indeterminate and refusal remain indeterminate. A one-source dry run on an old test positive verified all five artifacts and the local scoring path, but is pipeline validation only and contributes no confirmatory evidence.

2026-08-10: first confirmatory oracle attempt

Four new 2048x2048 Gemini outputs were generated after the local rule and strength were frozen, then registered as a 20-request batch. The manifest and all 20 artifacts passed hash verification before submission. The first source group produced these ordered results:

Role	Local tile/phase decision	Provider result
source	detected / detected	SynthID indeterminate; Google C2PA present
re-encode control	detected / detected	Google AI signal detected
aligned subtraction	detected / not detected	Google AI signal detected
shifted control	detected / detected	Google AI signal detected
orthogonal control	detected / detected	refused: verification quota exceeded

The quota response instructed the account to wait 24 hours. No adaptive retry was made, and the remaining 15 preregistered requests were not submitted. The batch cannot receive its preregistered aggregate verdict until complete ordered results exist.

Even incomplete, this attempt rejects the frozen removal recipe for the first temporal source: aligned subtraction crossed the phase threshold but remained above the tile threshold and did not clear the provider oracle. The remaining sources can still measure transfer and disagreement between the two local experts, but they cannot turn this first aligned result into a universal pixel-only removal success.

2026-08-10: positive-only runtime detector

Removal is deferred while the transferred tile signal is exposed as a bounded detector. src/remove_ai_watermarks/synthid_detector.py loads the frozen model as a bundled pickle-free runtime asset and returns detected, not_detected, or unsupported without resizing input. The direct API and detect-synthid CLI initially covered only native 2048x2048 images; identify consumes a positive match as high-confidence evidence but never treats a negative or unsupported result as proof of absence.

The operating point and model are unchanged from the locked experiment. The detector accepted all four post-freeze Gemini outputs, including the source whose frozen phase score missed, while retaining the prior 29-of-30 locked-test sensitivity and zero accepted validation/test negatives. The unresolved two external-source matches remain the reason a runtime positive names the carrier but does not attribute a provider.

2026-08-10: calibrated native-geometry extension

The frozen 16x16 template and threshold transfer across the common native geometries in the provider-positive evidence. Non-divisible image dimensions use direct modulo folding with per-cell sample counts; divisible dimensions, including 2048x2048, retain the original numerical path. All prior 2048x2048 records were replayed and matched exactly, including floating-point scores.

The runtime registry was expanded only after the unchanged threshold accepted none of 60,000 public COCO challenge views, 3,000 at each of 20 target geometries. This is a geometry extension of the same positive-only carrier expert, not an OpenAI pixel detector or a proprietary payload decoder.

2026-08-11: calibrated image-size range

The fixed template was then evaluated on every provider-positive image in the local evidence set rather than only its common geometries. It accepted 3,928 of 4,698 images across 757 exact geometries. Sensitivity separated by pixel count: 1,987 of 2,021 images at or above three megapixels crossed the threshold, while 1,940 of 2,672 images from one through two megapixels did. This establishes a carrier-family boundary, not universal SynthID recall: explicit C2PA watermark actions also occur below threshold, and three strong carriers use a different cyclic phase.

Two public-image geometry challenges tested whether geometry itself creates false matches. The first balanced 5,000 COCO images across all 757 observed geometries, with every geometry present in both development and final partitions; the maximum fixed score was 0.12549 and none crossed the unchanged 0.17357 threshold. The second transformed the same 5,000 source images at 256 generated geometries from one through 18 megapixels, covering every width/height remainder pair modulo 16. Its maximum was 0.16387 and again none crossed the threshold. Runtime support therefore uses that challenged pixel-count interval instead of an enumerated geometry registry, still without resizing.

A cyclic-registration branch was rejected after the public-image challenge. At a low threshold it recovered 130 fixed-phase misses but false-fired on tracked OpenAI and Adobe controls. Raising the threshold above all development controls recovered only 23 in-range provider positives while adding 84 cross-source matches beyond the fixed detector. Three strong shifted Google carriers remain a documented miss rather than weakening the runtime rule.

Top-level source labels were not treated as watermark-negative truth. Of 237 non-Google-labeled corpus images crossing the fixed threshold, 81 retained SynthID provenance. High-scoring Adobe manifests explicitly named Gemini Flash as the remote provider, and perceptual matching found direct Google derivatives under editing or relabeling categories. The remaining source labels cannot prove absence of the carrier, so they constrain provider attribution but do not justify raising the pixel threshold.

2026-08-11: spatial-resampling boundary

Arbitrary decoded geometry did not imply arbitrary resize robustness. A stratified sample of 80 fixed-positive provider images was rescaled in memory at 0.5, 0.625, 0.75, 0.875, 1.125, 1.25, and 1.5. The unchanged 16x16 detector accepted none of the 80 images at every nonidentity scale. Restoring each view to its source dimensions recovered 58-80, which isolates the failure to the scaled sampling lattice rather than simple information destruction.

Scaling the template and modulo fold to the corresponding integer periods 8, 10, 12, 14, 18, 20, and 24 recovered 73-80 of the discovery images at the native threshold. Searching all eight periods raised control scores, so a single threshold was frozen above the maximum of 3,000 deterministically resized COCO development views. It accepted none of 2,000 final controls and 672 of 800 source-disjoint provider positives, with 77-88 accepted at each scale. Six tracked OpenAI and Adobe controls remained below the frozen threshold.

This did not generalize to arbitrary scaling. Factors 0.8, 0.9, 1.1, 1.2, and 1.333 create noninteger carrier periods; rounding the period accepted zero of 80 images in the fixed-phase branch at every factor, and cyclic registration accepted only 1-16. Period-specific thresholds improved sensitivity but accepted five of the 2,000 final controls. The integer-period bank is therefore retained as a research result, not added to the runtime. A universal resize branch needs fractional-period registration or reliable reconstruction of the pre-resize geometry, followed by a new multiple-search calibration.

A discovery follow-up tested that fractional-period direction with the 30 strongest template harmonics. A continuous 7.5-24.5 spectral search placed the true period within 0.05 pixels among its top three candidates for 58 of 60 transformed positives. Searching nine neighboring integer reconstruction geometries recovered 44 of 60 at the native threshold; even the true source geometry recovered only 48, so resampling damage sets a separate upper bound. The full search then failed a frozen 250-development/150-final COCO challenge: the calibration threshold accepted two final controls. A preselected product of canonical correlation and spectral confidence increased positive acceptance from 49 to 52 of 60 at its development threshold but accepted the same two final controls. This fractional FFT baseline is rejected for runtime use; better period estimation alone does not solve multiple-search specificity.

2026-08-11: multiscale calibration and opt-in registered detector

The fractional search was revisited as a multiple-search calibration problem, not as period estimation alone. A source-disjoint challenge used 200 confirmed fixed-carrier positives, 40 independent sources at each scale 0.8, 0.9, 1.1, 1.2, and 1.333. Controls were 4,900 unique COCO images resized into the positive pixel-count domain, split into 2,900 development and 2,000 final items. The ten frozen candidates combined exact fractional canonical correlation, quadrant agreement, and a three-level scale-space score through mean or minimum rules. Pure box-counting dimensions, multifractal width, lacunarity, roughness, wavelet-energy slope, and a generic log-Mellin response had already failed the smaller pilot and were not promoted to the large challenge.

At the development-only threshold, the three-way mean accepted 177 of 200 positives versus 163 for canonical correlation alone, with 16 positives rescued and two lost in the paired decision. Each produced one final false positive, on different source images, so the development threshold was rejected. Recalibrating above all 4,900 COCO controls left 176 of 200 three-way positives and 160 of 200 canonical-only positives. This recalibration is exploratory rather than a new validation because the final controls had then been inspected.

A second source-disjoint corpus froze those thresholds before scoring 105 new positives across scales 0.5, 0.65, 0.8, 1.0, 1.2, 1.333, and 1.5, plus all 879 content-deduplicated Kodak and Picsum controls available locally. Natural 2816x1536 controls had a heavier null tail than upscaled COCO: the three-way mean accepted 82 of 105 positives and six controls at the COCO threshold, whereas canonical-only accepted 79 positives and ten controls. A threshold above all 5,779 controls retained 253 of the first 305 positives for the three-way mean. Four coarse pixel-count bins recovered only six more positives, so size normalization did not explain the main 0.5x miss.

At 0.5x, a fixed period-8 branch raised canonical correlation for seven of the 15 discovery positives but also won the unconstrained search on 629 of 879 controls. Searching it unconditionally merely raised the null threshold. The frozen remedy requires period-8 native correlation of at least 0.35 and takes the maximum of the old and period-8 three-way scores. It added four positives without an accepted control in the discovery corpus. On the earlier 4,900 controls, the frozen gate never opened and none crossed the final three-way threshold 0.2460603834083705.

The improvement itself was then tested on 50 additional source-disjoint 0.5x positives that influenced neither gate nor threshold. The ungated three-way score accepted 6; the frozen gate accepted 17, rescuing 11 and losing none. The exact paired sign test was p = 0.0009765625. These results support an opt-in scale-registered detector over 250,000 through 10,000,000 decoded pixels. The runtime now requires both sides to be at least 256 pixels because its frozen confirmation needs complete disjoint 256-pixel patch groups. Narrower inputs need a separately calibrated adaptive-patch expert. These results do not establish a universal SynthID decoder: the control maximum helped set the operating point, 33 of 50 new 0.5x positives still missed, and crop, strong codec changes, other carrier epochs, and provider attribution remain outside the claim. The fast native fold remains the default and the identify path.

That period-8 conclusion was overturned by a later symmetric negative test. Every one of the 11,506 known non-Google Spaces controls was downscaled by 0.5, matching the positive transformation; 11,273 outputs remained in the registered geometry range. The direct analytic 8-by-8 template was dominated by shared resize structure: 80.8% of its energy projected onto a 2-by-2 lattice and 94.0% onto a 4-by-4 lattice. The production period-8 gate opened on 73 controls, and all 73 crossed the registered threshold. They spanned hosted jobs, Firefly, China AIGC-labeled generators, Meta-tagged images, OpenAI, and local tools. No tested secondary condition separated the 11 rescued positives from those 73 controls. The explicit period-8 rescue is therefore rejected and not part of the runtime detector. The 0.5x rescue result is retained here as an incident record against asymmetric transformed-positive evaluation.

The same symmetric corpus also invalidated the main single-threshold registered rule: 68 of 11,273 supported controls crossed 0.2460603834083705. Fixed thresholds for nine recovered-period bins eliminated those errors while retaining 238 of 355 transformed positives, but one of 4,900 COCO controls then crossed through the third spectral candidate. Requiring the canonical winner to match the strongest spectral candidate removed that case and all 68 Spaces errors, with 239 of 355 positives retained. This version was not shipped: a fresh, content-deduplicated 1,000-image Open Images validation corpus produced six false positives.

All six Open Images errors concentrated their apparent match in one middle-frequency ring. Their correlations disappeared or changed sign in the two higher-frequency rings of the 16-by-16 carrier, unlike the accepted Gemini positives. Adding a frozen minimum correlation of 0.075 across those two rings removed the six development errors and retained 229 of 355 positives. A previously untouched reserve of 499 supported Open Images controls then produced five amplitude-gate candidates and zero final positives; their maximum high-band score was 0.004682175209264222.

The retained positives by scale were 0/65 at 0.5, 8/15 at 0.65, 53/55 at 0.8, 33/40 at 0.9, 13/15 at 1.0, 30/40 at 1.1, 38/55 at 1.2, 40/55 at 1.333, and 14/15 at 1.5. A physically transformed template recovered strong correlations for some period-8 positives, but 975 top-candidate Spaces controls fully overlapped every tested amplitude, spectral-band, and joint feature. No zero-control operating point retained a period-8 positive. The calibrated runtime therefore uses the top-candidate, period-bin, and high-band gates and makes no 0.5x detection claim.

An exploratory amplitude-times-high-band rescue later tested whether the two existing gates could trade evidence instead of both passing independently. A threshold frozen above all 341 top-1 development controls recovered 11 of the 126 current false negatives and accepted none of a separate 499-image Open Images reserve; the reserve maximum was 0.100583 against the 0.118168 threshold. The decisive symmetric challenge rejected it: 19 of all 11,506 Spaces controls after 0.5x resizing crossed the frozen threshold, with a maximum of 0.179989. The errors spanned OpenAI, Firefly, Microsoft, hosted jobs, and other generators. Raising the threshold post hoc above that maximum would recover only seven false negatives, none at scale 0.65 or 0.8. The joint rescue is rejected; the independent amplitude and high-band gates remain the measured operating point.

2026-08-13: public bypass corpus and inverse-carrier audit

The public cebeuq/Synthid-Bypass V2 comparison set supplied 12 external before/after pairs produced by a diffusion reconstruction workflow. The fixed runtime carrier accepted 10 of 12 before images and none of the 12 after images. The two misses were not assumed negative: Gemini Verify AI reported SynthID on before/03.webp and before/07.png, and no SynthID on both paired outputs. Outputs 01 through 06 were also oracle-negative; the verifier's daily limit stopped the same-session sweep at 08.

07.png is a low-SNR instance of the shipped positive direction. Its full-frame score was 0.1248368819; its signed RGB correlations were 0.073, 0.183, and 0.111, and all nine aligned region scores were positive. A proposed rescue at score >= 0.12 with Green dominance and spatial agreement separated it from the two highest transformed-COCO controls, but failed a fresh exact-1024 Open Images challenge. A 2009 natural photograph scored 0.2123730698, had stronger Green dominance and nine-region agreement, and Gemini returned SynthID not detected. The downloaded bytes matched a fresh fetch from the official Open Images validation bucket. This is a confirmed runtime false positive, not a provider-label inference.

The failure is broader than that one JPEG. Symmetrically resizing the 24 Kodak PNG controls to 1536 by 2752 with Lanczos produced fixed scores as high as 0.3768339306. Those synthetic views were not used as independent clean samples, but they prove that rational resize lattices can create a stronger signed 16-by-16 match than the runtime threshold. The existing 5,000-view observed-geometry and 5,000-view generated-geometry COCO challenges remain valid for their exact source-and-transform assignments; they do not justify a universal transformation claim. No threshold relaxation for 07 is allowed, and the current direct detector needs a fresh source-and-transform-grouped specificity epoch before its support range can be broadened.

03.webp tested a different hypothesis. The two content-group-separated alternate-Gemini experts both rejected it. Its shipped-template score was instead -0.3263447029, all three channel scores were below -0.28, and every one of nine aligned regions was negative. The paired oracle-negative output weakened to -0.1531420136. Neither 5,000-view COCO geometry challenge nor the fresh Open Images blocks contained a score at or below -0.20, making the opposite-polarity response a plausible additional codeword or provider state. It is not specific enough to ship: 40 of 11,506 heterogeneous non-Google-labeled AI controls crossed -0.20, with examples from OpenAI, Microsoft, multiple other generators, and unassigned made-with-AI files. An inverse-template expert therefore remains a research hypothesis until matching-provider oracle labels and a new untouched natural and AI control split exist.

A follow-up decomposed the two high-frequency rings instead of taking their minimum. The 19 joint-rule false positives were broad-band, with median min(ring) / max(ring) of 0.84, while the registered false negatives had a median of 0.27. A development-selected rescue therefore multiplied normalized amplitude by the stronger ring and required ring balance at or below 0.5. A threshold above 341 development controls recovered 11 of 65 top-1 false negatives and accepted none of the 499-image Open Images reserve. The full symmetric challenge again rejected it: 13 of 11,506 resized Spaces controls passed, with a maximum score of 0.694818 against the 0.308683 threshold. Strong single-ring lattices occurred in Firefly, OpenAI, hosted jobs, China-labeled generators, Canva, and other controls. Radial imbalance is therefore another general renderer or resize feature, not a SynthID-specific rescue.

A multi-codeword hypothesis then projected the shared 16-by-16 template out of each normalized residual tile and fitted source-disjoint spherical codebooks to the remaining 768-dimensional vectors. Five prototypes selected on 300 positives retained 366 of 600 held-out positives. Three prototypes were highly reproducible across an independent 300-positive fit, with matched correlations of 0.88 through 0.95, while the additional states were less stable. A threshold above 2,500 development natural-image controls accepted none of a separate 2,500-image reserve and selected 225 of 770 metadata-labeled Google images that the shared-template threshold missed. The broader challenge decisively rejected the feature: 828 of 11,506 metadata-labeled non-Google Spaces controls crossed the frozen threshold. Errors occurred across OpenAI, Microsoft, Firefly, China-labeled generators, hosted jobs, Stable Diffusion, and other sources. A post-hoc intersection above every control that also passed the registered min(R, B) gate retained 138 Google candidates, but none passed that independent carrier gate. The residual clusters are reproducible generator or resampling structure, not a specific additional SynthID codebook, and are excluded from runtime.

A related spatial-payload pilot stopped averaging all 16-by-16 blocks and measured carrier and codebook responses block by block. The positive carrier response was spatially coherent: median adjacent-block correlation was 0.61 on both axes, compared with 0.20 and 0.24 in a heterogeneous Spaces sample and 0.04 and 0.06 among codebook false positives. The development-selected rescue was instead the standard deviation of the best codebook response across blocks. It retained 11 of 60 reserve positives and none of 180 reserve controls, but accepted six of 1,000 fresh heterogeneous controls. Those errors included OpenAI, hosted, and China-labeled generators; raising the threshold above their maximum retained only four of the 120 pilot positives. Per-block carrier or codeword coherence is therefore diagnostic but not a usable rescue.

A geometry-stratified follow-up found a second reproducible native 16-by-16 carrier in the fixed-template misses. Templates fitted independently from two halves of the 1024-by-1024 misses correlated at 0.869, and two independent fits from the non-1024 misses correlated at 0.956. The latter correlated 0.82-0.84 with the 1024 template but only about 0.28 with the shipped carrier. This rules out a single-resolution averaging accident and supports a distinct carrier direction or state.

The initial 1024 template retained 8 of 91 held-out misses and accepted none of 81 exact-geometry reserve controls or 500 natural-image reserve controls. Seven of those eight positives survived JPEG 95, six survived JPEG 85 and 70, and three survived a 0.75x resize round trip; WebP 95 retained none. A threshold frozen above the exact-geometry development controls accepted 25 of 11,506 heterogeneous controls when applied at arbitrary native geometries, so an unrestricted second native template is rejected. Canonicalizing every control to 1024 by 1024 eliminated all 11,506 errors, and the same path accepted only the native 1024-by-1024 Google misses. The signal has a fixed native period, not a size-normalized frequency.

A stricter chronological fit used only 1024-by-1024 misses through mid-June and tested July misses. Its template correlated 0.839 with the independently fitted July template, retained 5 of 95 temporal misses, and accepted none of 81 reserve exact-geometry controls. At that frozen threshold, canonicalization accepted 18 of all 770 fixed-template Google misses and none of all 11,506 heterogeneous controls; every accepted miss was natively 1024 by 1024. Requiring agreement between two independently fitted early templates retained three temporal misses and no reserve controls. This is credible evidence for a second carrier direction within the Google cohort, but the positive and exact-geometry control counts are still too small for a runtime operating point. The later external natural-image challenge below further shows that this repeatability is not yet specific enough for blind detection.

Projecting the shipped carrier out of every normalized tile did not remove the second direction. The early orthogonal template was numerically orthogonal to the shipped template, correlated 0.834 with an independently fitted July orthogonal template, and retained the same 5 of 95 temporal misses with none of 81 reserve exact-geometry controls. Its threshold was frozen above the development controls before the decisive 1024-by-1024 canonical challenge; none of all 11,506 heterogeneous controls crossed it, and the maximum remained just below the frozen boundary. The second direction is therefore not a weak projection of the shipped carrier. This strengthens the research finding but does not enlarge the small temporal-positive or exact-geometry control reserves required for runtime.

Mapping both frozen scores over all 4,698 Google-labeled images changed the interpretation from an epoch replacement to coexisting carrier states. The map contained 3,825 first-only, 69 second-only, 103 both, and 701 neither cases; the two scores correlated at -0.563. Both directions occurred throughout the collection interval and across multiple native geometries, while the presence of both in 103 images rules out a simple dated switch from one template to the other. The second native threshold remains exploratory outside exact 1024-by-1024 inputs because 8 of 11,506 heterogeneous native controls crossed the orthogonal version's threshold; only the canonical exact-geometry challenge retained the zero-of-11,506 result.

Freezing a stricter native threshold above the maximum of 5,753 alternating development controls left zero crossings in the other 5,753 controls and 26 Google images missed by the first carrier above threshold. That retrospective split is promising but not a certified operating point: the corpus had already been inspected while developing the second carrier, so a fresh source- and time-disjoint challenge is still required.

An external natural-image challenge then rejected that strict native threshold. It accepted none of 1,500 unique Open Images controls and none of 12 Wikimedia controls, but one of 637 content-deduplicated Picsum controls. The original Picsum directory contained 1,000 filenames but only 637 unique hashes; the one crossing appeared twice under different ids and is counted once. Its RGB and quadrant contributions were highly uneven, unlike the small temporal-positive set, but that post-error observation cannot retroactively define a gate. The strict native threshold remains excluded from runtime.

The original exact-1024 threshold also failed on those external controls: four of 1,500 Open Images and four of 637 unique Picsum images crossed it, including native 1024-by-1024 cases. The earlier zero-of-11,506 canonical result was therefore specific to that heterogeneous control composition, not evidence that exact geometry alone makes the carrier specific. Requiring agreement between the two independently fitted early templates did not close the gap; it retained three temporal positives but accepted one Open Images and two unique Picsum controls. The second direction remains a reproducible Google-cohort signal, not a blind detector.

Requiring the orthogonal score to survive a JPEG-95 round trip was not a specificity gate either. It retained four of the five temporal positives but also retained all four Open Images and all four unique Picsum crossings at the original threshold. The natural-image confound is itself transport-stable. Together with the failed low-rank, two-template, channel-balance, quadrant, and two-carrier-plane variants, this exhausts mean-template consensus for the second direction on the current data.

The orthogonal carrier is not a cyclic shift or color permutation of the shipped template. Its best absolute cyclic correlation was 0.212, Fourier magnitude correlation was 0.480, and weighted phase lock was 0.200. Dominant frequencies were axial, led by (0, +/-8), (0, +/-3), and (0, +/-5) rather than the first direction's broader structure. A horizontal-axis-only ablation retained all five temporal detections and no reserve exact-geometry control, but then accepted three of the 11,506 canonicalized heterogeneous controls. The errors included an OpenAI-labeled image and two watermark-remover outputs. The full orthogonal template is more specific than its strongest axial component.

A final low-rank variant tested whether that direction contains several payload states rather than one mean carrier. SVD bases of ranks 1, 2, 3, 4, 6, 8, 12, and 16 were fitted only on early 1024-by-1024 misses. Rank was selected on a separate June interval and a separate control third, then reported on July misses and the final control third. The selected rank 12 retained 2 of 38 validation misses and 4 of 95 test misses, with no control error. Rank 2 happened to retain 6 test misses but only one validation miss and therefore could not be selected without post-test tuning. A multi-state subspace did not improve the evidence-grade operating point.

Deflation then tested whether a third stable linear carrier remained. After projecting both the shipped and second orthogonal templates out of every normalized tile, early and July residual averages correlated only 0.0003. The early residual template retained none of 95 temporal misses and none of 81 reserve exact-geometry controls. Within this 1024-by-1024 temporal corpus, the linear periodic model therefore supports two reproducible carrier directions, not an open-ended sequence of mean templates.

Treating those directions as a single two-dimensional carrier plane did not provide a safer weak-signal rescue. A positive-quadrant Euclidean norm used the two independently frozen carrier thresholds as axis scales, then froze its own boundary above 81 exact-geometry development controls. It rescued none of the Google images below both component thresholds and accepted one of 81 reserve controls. The two directions must retain separate operating points; combined subthreshold energy is rejected.

The five temporal positives above the original second-carrier threshold also had unusually balanced normalized R, G, and B contributions: their worst channel ranged from 0.297 to 0.320, while seven of eight native control errors were below 0.297 and were commonly blue-heavy with a weak green contribution. A clean chronological validation did not justify turning that observation into a gate. A template fitted on half of the early positives produced no accepted examples in the other early half at the threshold frozen from development controls, despite accepting six July examples and no reserve control. The channel-balance rule is therefore a post-test hypothesis, not detector logic.

A phase-only follow-up removed observed Fourier magnitude and averaged phase agreement using only the early template's fixed spectral-energy weights. The phase threshold alone retained 2 of 31 early validation misses and 10 of 95 July misses, but accepted one of 81 reserve exact-geometry controls. Requiring both phase agreement and the independently calibrated linear projection removed that control error and all external natural-image errors. The frozen conjunction retained 1 of 31 early validation misses and 3 of 95 July misses, accepted none of 162 heterogeneous exact-1024 controls, none of 1,500 unique Open Images, none of 637 unique Picsum images, and none of 12 Wikimedia images. Applied outside its declared geometry it accepted 3 of 11,506 heterogeneous native controls, all at non-1024 sizes. This is the first independently challenged phase-and-amplitude candidate for the second direction, but three temporal positives are not enough to ship a runtime branch.

Across all 443 exact-1024 Google images, the frozen conjunction found 11 that the shipped carrier missed and 16 that both branches accepted; 244 were found only by the shipped carrier and 172 by neither. The candidate therefore adds signal rather than merely confirming the first direction, but its incremental native recall is still small.

Swapping the two early fit halves and recalibrating both thresholds retained 4 of 31 opposite-half validation misses and 4 of 95 July misses, again with none of 81 reserve exact-geometry controls passing the conjunction. This supports fit-split signal reproducibility, but the swapped operating point accepted one of 92 unique exact-1024 Open Images controls in its external challenge; none of 637 unique exact-1024 Picsum controls crossed. A post-hoc consensus of the two fit halves would require another untouched control set, so it is not selected from these exposed results. The candidate remains research-only.

An initial two-expert rule required both independently fitted templates to pass their own separately calibrated linear and phase thresholds. The next nonoverlapping Open Images validation-id block was downloaded without scoring; 2,000 candidates yielded 100 unique native 1024-by-1024 controls. The first expert accepted none, the second accepted one, and their conjunction accepted none. That rule appeared to add 5 detections across the 443 exact-1024 Google files, including 2 of 95 July fixed-carrier misses.

A subsequent content-group audit invalidated those recall numbers. The 443 files contained three exact decoded-pixel duplicate pairs and ten perceptual-hash groups at a conservative 16-of-256-bit screening distance. One early near-duplicate pair had been split between the two expert fits, so the experts were not content-independent. After allocating whole groups to one side, selecting one representative per group, and recalibrating both experts from the same untouched development controls produced templates correlated at 0.720, and the consensus retained only 6 of 443 Google files. It added 2 files missed by the shipped carrier and retained only 1 of 95 July misses. It still accepted no conjunction among 81 reserve exact-geometry controls, 92 exact-1024 Open Images, 637 exact-1024 Picsum images, or the 100 new Open Images controls. The corrected result is specific but far too small to ship.

Removing the exact-geometry restriction invalidated even that corrected consensus. Across all 11,506 heterogeneous native controls, the group-separated experts accepted 5 and 6 respectively, and their conjunction accepted 2. Both consensus errors were at non-1024 geometries. The result is explicitly an exact-1024 research candidate and does not supply a universal-size detector.

A post-hoc margin above both arbitrary-geometry collisions retained the two corrected incremental exact-1024 positives. Applying that frozen margin to all 4,698 Google files accepted 28 and added 16 files missed by the shipped carrier across 11 native geometries. A full-corpus perceptual audit assigned those 16 files to 16 separate content groups, none containing a shipped-carrier-positive variant. The same margin accepted none of 1,500 Open Images, 637 unique Picsum, or 12 Wikimedia native views. This is the first concrete all-size hypothesis for the second direction, but the margin was selected after all 11,506 heterogeneous controls were exposed and the natural corpora had already challenged related experts. It requires a new content-deduplicated AI-control corpus before it can be treated as validation or runtime logic.

The next source-independent model-cohort challenge sharply bounded that hypothesis. With no exact-byte overlap against the Spaces inventory, the original group-separated conjunction accepted 2 of 589 public Gemini 3.1 Flash Image Preview images, none of 520 Nano Banana Pro Preview, and none of 280 DALL-E 3 images. The two Gemini hits were visually distinct diverse images at 1408 by 768. Each Google cohort also included solid-color and gradient probes over several native geometries. The post-hoc 1.033 strict margin rejected every cohort image: Gemini reached 1.011, Nano Banana 0.941, and DALL-E 3 0.514. The base rule therefore has weak transfer to a current Gemini cohort but still collides with 2 of 11,506 arbitrary-size controls; the strict rule removes both the controls and the current-Gemini transfer only through a post-test threshold. Neither is a validated universal Google-model or cross-provider SynthID detector. The 16 Spaces all-size hits and two public Gemini base hits are useful hard positives for epoch analysis, not justification for implementation.

The near-duplicate audit also provided a small mechanistic diagnostic. In all three perceptually matched pairs where only one variant crossed the shipped carrier, subtracting the shipped variant from the non-shipped variant reduced the shipped-template direction and increased both independently refitted second-template directions. The normalized mean paired difference correlated -0.568 with the shipped template and 0.252 and 0.264 with the two second templates. Three pairs are insufficient for inference; the two-sided sign-test result is 0.25. The paired result supports a carrier-state interpretation but is not detector validation.

All three temporal detections from the original first-expert conjunction survived JPEG round trips at qualities 95, 85, and 70. None survived WebP 95 or a 0.75x down-and-up resize round trip. A symmetric transform challenge prevented promoting that positive retention to a JPEG claim: one of 162 heterogeneous exact-1024 controls crossed the unchanged rule after JPEG 85. The 92 exact-1024 Open Images and 637 unique exact-1024 Picsum controls stayed below threshold in every view. The candidate was native-only; fixed-lattice JPEG retention is diagnostic, not a certified transport operating point. The transform result applies only to that first-expert rule. The corrected group-separated experts have not been transport-calibrated, so it supplies no robustness claim for them.

The corrected carrier was then tested causally. The normalized sum of the two content-group-separated expert directions correlated 0.927 with each expert. Subtracting it from all 16 strict incremental hits cleared every linear and phase component on every image; the same amplitudes with a one-pixel cyclic shift cleared none. Selected amplitudes were 5.0 through 7.75 integer levels, with median 6.5. Fidelity ranged from 58.92 through 69.97 dB PSNR and 0.99961 through 0.99996 SSIM. The carrier itself was transform-fragile: only one of 16 sources remained strict after JPEG 95 or 90, and the aligned edit cleared that one in both views while the shifted edit did not. No source remained strict after JPEG 85, WebP 95, or a 0.75x resize round trip.

A joint ablation then started from all 28 strict second-carrier hits. It first suppressed the shipped carrier to the frozen -0.25 target where required, then suppressed the second direction until all four expert components were below their base thresholds. The original cohort contained 12 first-carrier and 28 strict second-carrier detections; the aligned candidates contained zero of either. The shifted controls left one first-carrier and seven strict second-carrier detections. Median fidelity was 61.69 dB PSNR and 0.99984 SSIM; the minima were 58.15 dB and 0.99910. The two edits did not reactivate one another. Together with the failed third-carrier fit, the current linear native 16x16 hypothesis is locally exhausted as two jointly controllable states. These are detector-score interventions, not Google-oracle removals.

2026-08-12: Registered color and phase-lock challenge

A crop-specific research branch tested whether the recovered 16x16 carrier is better represented in a perceptual color space or a shift-tolerant directional transform. Every branch reused the direct RGB period-and-phase registration, fitted on the first 30 positives, selected on the next 20 positives and 200 controls, and reported the remaining 50 positives and 299 controls separately. Only records whose recovered period was 15.5 through 16.5 were eligible.

The frozen RGB template's channel norms were 0.887:1.000:0.930 for R:G:B, including 0.903:1.000:0.916 above FFT radius 4.5. An SVD assigned 79.80% of template energy to a nearly equal-channel component, 18.39% to a green-vs-magenta component, and 1.81% to a red-vs-blue component. The common component had strong diagonal energy, while 98.6% of the coarse green-opponent orientation energy was axis-aligned. This measured carrier does not support a fixed 0.85:1.00:0.70 channel rule or an exclusively diagonal decoder.

Nonlinear cube-root LMS coordinates from the OKLab transform exposed the signal most strongly in the direct color-fold comparison, but the development-selected full-vector candidate retained 32 of 50 final positives and accepted six of 299 controls. This is evidence that nonlinear color projections can improve carrier SNR, not evidence that the deployed embedder uses Lab or OKLab internally.

A three-level DTCWT probe retained all six orientations rather than only the +/-45 degree pair. Its development-selected near-vertical green energy candidate retained 37 of 50 final positives and accepted eight controls. The green-opponent near-vertical candidate retained 39 positives and accepted one control. Separating energy, normalized complex correlation, and phase-only coherence changed the development choice to all-orientation green-opponent phase coherence, which retained 37 positives and accepted four controls. The directional representation therefore exposes real carrier structure but fails the low-false-positive gate.

An explicit FFT probe then froze 8, 16, 32, or 64 carrier bins from the fit positives and compared signed cosine phase coherence with the proposed abs(cos(delta_phase)) > 0.90 rule. The absolute rule never produced a usable operating point; treating phase and antiphase as equivalent increased the null overlap. With 64 bins, the development-selected green-channel signed cosine retained 26 of 50 final positives and initially accepted none of 299 controls. That result was underpowered because only 39 of 499 controls selected the eligible period-16 branch.

The green phase candidate and its threshold were frozen before a full period-and-phase search over 1,000 COCO controls. It accepted five controls. Among the 129 eligible controls, the maximum score was 0.6910271977636098, above the frozen 0.5443545641095117 threshold. Raising the threshold above that maximum retained only eight of 100 positives. Explicit registered phase lock is therefore a useful diagnostic of this carrier family but is rejected as a standalone detector. Because it shares the RGB registration stage, it is also not independent confirmation of the carrier.

2026-08-12: Green phase perturbation and blind SWT removal

A second removal probe tested the specific claim that changing only phase at the recovered Green-channel carrier bins can preserve spectral magnitude and image fidelity while defeating detection. It used the 100-image public Gemini set at native period 16 and modified conjugate FFT pairs together so the inverse remained real. Before pixel quantization, the maximum absolute change in FFT magnitude was 5.24e-10 for the 32-pair inversion. After clipping and rounding to 8-bit pixels, the maximum magnitude change relative to the largest source coefficient was 0.175%.

The native periodic detector accepted 59 of the 100 sources. Inverting the 32 strongest unique Green-channel carrier pairs left one of those 59 positive. Across all 100 images, PSNR was at least 43.36 dB and SSIM at least 0.99048; the medians were 52.79 dB and 0.99739. Median signed phase coherence moved from 0.485 to -0.371 and the median full RGB tile score moved from 0.230 to 0.041. This reproduces the proposed fidelity and local-score effect against the recovered periodic detector.

It does not establish SynthID removal. The current test targets the same carrier used by the local score, and earlier wider complex-spectrum and phase-projection candidates already cleared local phase models at 51.94-56.03 dB while remaining detected in healthy Google verifier sessions. The 32-pair Green candidate is therefore a preregistered oracle candidate, not a remover. The official SynthID-Image description also identifies a post-hoc neural, content-dependent encoder-decoder rather than a fixed Fourier codebook; it does not document phase bins as its decision surface.

The preregistered Google-oracle challenge then used three of the strongest local positives. Each fixed group contained the original, its 32-pair Green phase inversion, and an RMS-matched random-noise sham. All three phase candidates fell below the local 0.17357069773071196 threshold, with PSNR from 55.41 to 57.86 dB and SSIM from 0.99709 to 0.99807. The three shams remained locally positive. Nine files were submitted to Gemini's Verify with SynthID tool in separate chats with the unchanged prompt. Eight verdicts persisted and all eight reported Google AI signals. This includes all three phase candidates, both recoverable source controls, and all three shams. The second source request was submitted, but its chat did not persist and no verdict could be recovered; it was not retried outside the fixed upload budget.

This falsifies the operative removal claim: Green-carrier phase inversion clears the local periodic correlate on three of three images but clears the Google oracle on zero of three. It is evidence that the local template observes one stable component or correlate of SynthID, not that these FFT coefficients form the proprietary decoder's sufficient decision surface. Do not productize the Green phase inversion or report its local score change as SynthID removal.

The three oracle-confirmed phase candidates were then treated as adversarial hard positives for detector analysis, without using them to relax the existing runtime thresholds. Both runtime modes missed all three. Their native scores were 0.126-0.160 against the 0.174 threshold; their registered decision scores were 0, 0, and 0.553 against the threshold of 1. A phase-invariant magnitude profile was therefore tested as a possible rescue. The development-selected all-bin RGB profile retained 39 of 50 final positives but accepted one of 299 final Open Images controls and still missed all three hard positives.

A frozen-template, top-32 Blue-magnitude profile did accept all three hard positives and initially accepted none of 499 Open Images controls, but this was a post-hoc hard-positive-selected branch with low ordinary sensitivity: 17 of 70 development and final positives. Its frozen threshold accepted 39 of 1,000 fresh COCO controls. Combining it with a native-score floor of 0.10 happened to accept none of all 3,000 COCO controls while retaining the three hard positives, but failed the next independent challenge: 90 of 11,506 metadata-labeled non-Google controls passed both gates. Accepted controls included Adobe Firefly, local diffusion, Samsung, Microsoft, and several other generator families. The phase-invariant magnitude rescue is therefore rejected as a provider-specific detector; it measures a shared periodic rendering family.

Channel separation exposed a narrower attack-specific rescue. The Green-only phase inversion leaves Red and Blue pixels unchanged, but the runtime RGB score normalizes all three channels jointly. On the three oracle-confirmed hard positives, separate signed correlations remained 0.781-0.804 in Red and 0.798-0.825 in Blue while Green moved to -0.400 through -0.645. A Red threshold of 0.3770237083693908, frozen above 200 Open Images calibration controls, retained 16 of 70 development and final ordinary positives and accepted none of the remaining 299 Open Images controls. It also accepted all three oracle hard positives and none of 3,000 fresh COCO controls.

Red correlation alone remained a generic periodic-renderer correlate: it accepted 96 of 11,506 metadata-labeled non-Google controls. Every one of those 96 had positive Green correlation. The attack-specific conjunction of Red at or above the frozen threshold and negative Green accepted none of 15,005 total controls across three independent corpora, while retaining all three oracle-confirmed phase candidates. Across the 100-image synthetic inversion probe, it recovered 19 of the 58 locally detected sources that the attack had moved below the RGB threshold. One additional attacked image still passed the existing detector.

This channel-discordance rule is the first zero-control rescue for the oracle-confirmed hard positives, but it is not added to runtime. Its form was motivated after inspecting the attack, only three examples carry external oracle labels, and its synthetic rescue sensitivity is 19 of 58.

A follow-up challenge registered that frozen rule over all 256 modulo-16 phase offsets and six fixed carrier periods: 8, 10.4, 12, 16, 20, and 24 pixels. The Red-only form accepted one of 3,000 COCO controls during phase registration; period search added no further acceptance. The accepted control scored 0.395 in Red but only 0.267 in Blue. This rejects Red-only registered rescue at the existing threshold.

The previously frozen min(Red, Blue) >= 0.3770237083693908 candidate, still combined with negative Green, accepted none of the same 3,000 controls under phase-only or phase-and-period registration. Its maximum registered control score was 0.342. A symmetric transform stress test then evaluated all twelve identity, crop, JPEG, and resize views for 100 fixed Open Images controls. None of the 1,200 inputs was accepted; the maximum was 0.332.

The stricter rule retained all three oracle-confirmed hard positives at native scale and after offsets of 1x1 and 7x11 pixels or a centered 5% crop. Period registration retained all three after 0.65x, 0.75x, 1.25x, and 1.5x resizing. It retained none after 0.5x resizing or JPEG round trips at qualities 95, 85, and 70. Every source and RMS-noise sham remained negative in all twelve views. The small-control zero was not sufficient for the complete period bank. A later five-period search over all 11,506 metadata-labeled non-Google controls accepted 16 at the original 0.377 threshold: two at period 20 and 14 at period 24. The maximum min(Red, Blue) score was 0.435813. This rejects one common threshold across periods 10.4 through 24.

A conservative high-score branch was then frozen strictly above that measured maximum. It retained all three hard positives at periods 12, 16, 20, and 24, corresponding to 0.75x through 1.5x resizing, identity, and the three crop views. A symmetric full-corpus resize challenge accepted none of 11,506 controls at 0.75x, 1.25x, or 1.5x; the respective maxima were 0.431352, 0.393949, and 0.372036. The 0.65x path needed separate treatment. Its native period-10.4 null maximum was 0.361367, below all three hard positives, but a symmetric 0.65x resize round trip raised one control to 0.397396, above the weakest hard positive at 0.394911.

That collision motivated an exploratory low-period channel-ratio gate: min(Red, Blue) >= 0.3770237083693908, negative Green, and Blue >= 1.1 * Red, only at period 10.4. It retained all three 0.65x hard positives. Among all 11,506 controls, the largest score satisfying the ratio gate was 0.355165 natively and 0.354955 after the symmetric 0.65x round trip, so neither challenge accepted a control. The combined exploratory rule also retained all three hard positives in the identity, three crop, 0.65x, 0.75x, 1.25x, and 1.5x views, while every corresponding source and RMS-noise sham remained negative.

Applying the combined rule without further tuning to the 100-image synthetic Green-inversion cohort exposed its low sensitivity. Among the 58 attacks that moved a natively detected source below the native detector threshold, the rule recovered 4 at 0.65x, 5 at 0.75x, 16 at identity, 11 at 1.25x, and 10 at 1.5x. The three oracle source images are members of that cohort; only one additional synthetic source passed the 0.65x branch. The rule is therefore a high-specificity rescue for a strong residual R/B carrier, not a general solution to Green-phase perturbation.

A source-split Fisher projection then tested whether a learned three-channel direction could improve that sensitivity. Red, Green, and Blue correlations were standardized separately by scale, and the projection was fitted from 34 training attacks and 5,719 training controls. Per-scale thresholds were frozen above every training-control score. They accepted 11 of 23,148 views from the 5,787 test controls. On the seven source-held-out attacks, the projection recovered 0 at 0.65x, 0 at 0.75x, 2 at 1.25x, and 2 at 1.5x. It is both less specific and less sensitive than the explicit rule and is rejected.

Spatial consistency at 0.65x was also insufficient. At the globally selected period-10.4 phase, a 2x2 and 4x4 aligned grid measured local R/B strength, negative-G coverage, and Blue / Red coverage on all 537 controls in the global-score tail at or above 0.20 and on all 58 synthetic native-detector misses. Every individual feature overlapped the control tail. Within a small fixed grid of global floors and integer coverage counts, the best zero-control conjunction recovered four attacks, no more than the simpler global ratio rule. Lowering the global floor to 0.30 recovered five but accepted two controls. The spatial branch is rejected as non-improving.

This establishes bounded transform behavior for the attack diagnostic, not a universal detector. The 0.65x ratio was selected after seeing the full-corpus collision, the conservative threshold was derived from an exposed control corpus, the period bank is discrete, only three recovered images have oracle labels, and JPEG sensitivity is zero in this probe. The next valid gate remains new oracle-confirmed Green-perturbed positives plus a preregistered untouched control corpus. The branch stays outside runtime until that gate passes.

An exploratory channel-contrast extension then tested whether the magnitude of the negative Green response could recover the failed JPEG and 0.5x views. The metrics were fixed as min(Red, Blue) - Green and mean(Red, Blue) - Green; 50 Open Images controls across all twelve transforms set thresholds before the remaining 50 controls were scored. Calibration over 600 views required thresholds of 0.655 and 0.695. Neither metric accepted an evaluation control, but both still recovered zero of three JPEG cases at every quality and only one of three 0.5x cases. A relaxed R/B-only threshold produced the same result. Negative-G magnitude therefore adds no useful transform coverage at the measured low-false-positive boundary and is rejected.

Codec-specific templates were then fitted from 50 other public positives after JPEG 95, 85, and 70 round trips, excluding all three oracle hard-positive sources. Neither a quality-matched template nor a three-quality bank recovered any of the nine JPEG hard-positive views. They accepted respectively three and two of 150 evaluation-control views. The templates primarily captured the JPEG 8x8 grid and were rejected.

Filtering the unchanged 16x16 template to radial frequency 0.5 through 4.5 exposed a real JPEG-surviving component. At a threshold above the first 300 Open Images JPEG views it retained all three hard positives at qualities 95, 85, and 70. The next frozen challenge rejected it: phase search accepted 35 native, 50 JPEG-95, 48 JPEG-85, and 46 JPEG-70 views among 3,000 COCO images. Cross-band phase agreement did not separate the nine hard-positive views from those 179 false positives.

A spatial gate required negative Green in all four quadrants and a median quadrant R/B score above the COCO false-positive maximum. It retained all three hard positives at JPEG 95 and 85 but none at JPEG 70. On 11,506 metadata-labeled non-Google controls it nevertheless accepted 35 native, 66 JPEG-95, and 88 JPEG-85 views. The errors included genuine camera images and multiple unrelated generators. Scoring only the unchanged phase also failed, accepting 38 native and 210 JPEG views among 3,000 COCO images; its maximum control score of 0.673 exceeded the 0.464-0.609 hard-positive range. The JPEG-stable band is a general low-frequency or codec lattice, not a SynthID-specific rescue, and all of these branches are rejected.

The 0.5x boundary was challenged separately with a physically reduced 8x8 template. Projecting the dominant 2x2 lattice out of both observation and template before scoring R/B with negative Green removed the known nuisance but not the overlap. Across all 11,506 symmetrically downscaled controls, the zero-control threshold was 0.396. The three oracle hard positives scored only 0.277-0.288. Channel separation therefore cannot rescue the rejected period-8 branch.

An untouched 2026-08-07 through 2026-08-10 temporal cohort then challenged the runtime template and thresholds. Cohorts were recomputed from explicit provenance because a provider-only split incorrectly placed mixed Google/OpenAI lineage in the negative set. The frozen native detector accepted 183 of 252 files with explicit Google LLC SynthID provenance and 3 of 488 with explicit OpenAI-only SynthID provenance. Scale registration accepted 130 and 1, respectively. Among 170 files with no SynthID provenance assertion, native and registered modes accepted eight and two. These operational errors are not provider-oracle negatives and cannot establish the true false-positive rate.

The native detector remained positive on 153 of 252 explicit-Google images after a symmetric 0.75x resize round trip, 66 after JPEG-95, and 54 after WebP-95. A 5% center-crop round trip retained none. This fresh temporal result confirms a real, partly codec-tolerant Google pixel signal across many stored sizes, while rejecting the stronger claim that the fixed-phase runtime score is a universal SynthID decoder. The post-hoc strict second carrier did not generalize: it accepted only 2 of 234 Google-platform positives and none of 689 provider-split controls. Applying the already frozen scale-registration path after the crop recovered only 8 of 252 explicit-Google files and accepted none of 488 explicit OpenAI-only files. The existing scale search does not solve the crop phase boundary.

A diagnostic joint-scale-and-phase probe then reversed the known 0.95 crop scale and scored the carrier at the cyclic phase predicted from the removed border. It reached 0.859 AUC for explicit Google provenance against all other temporal cohorts. A post-hoc threshold above all 671 other rows retained 37 of 252 explicit-Google rows. Exhaustively searching all 256 tile phases was weaker at 0.839 AUC and retained 41 rows above its own zero-control threshold. The fixed geometric phase supplies a plausible crop mechanism and a candidate for the next preregistered temporal challenge, not a calibrated runtime mode: both the transform and reported operating point have now observed this cohort.

A blind three-level db2 SWT probe then soft-thresholded only the Green horizontal, vertical, or diagonal detail bands at 0.25, 0.5, 1, and 2 times the band's robust noise estimate. Diagonal HH thresholding left all 59 original positives detected at every strength. Vertical thresholding cleared four at strength 1 and five at strength 2, but the respective minimum fidelity fell to 32.57 dB and 0.96971 SSIM, then 28.52 dB and 0.93646. At strength 1, vertical score reduction exceeded diagonal reduction on 97 of 100 paired images (p = 2.63e-25, exact two-sided sign test). Blind HH soft-thresholding is therefore rejected; the stronger vertical response is consistent with the separately measured axis-aligned green-opponent carrier component.

2026-08-13: patent architecture and amplitude confound audit

The official paper intentionally leaves the neural network architecture unspecified, but a related DeepMind patent family with overlapping authors gives more specific architectural evidence. Its image example constructs a content-dependent residual as x' = x + g(x), describes a U-Net-like encoder and convolutional decoder, injects a message or secret into intermediate layers, and trains clean and watermarked pairs under the same sampled transformations. It also permits a bank of paired networks in which one decoder is deliberately unable to recognize another pair's mark. These patent alternatives are not proof of the exact production implementation, but they invalidate the working assumption that SynthID must reduce to one provider-independent, fixed spread-spectrum key. Multiple Google states and a distinct OpenAI carrier are compatible with one technology family.

This changes the role of the shipped 16-by-16 template. It is a validated linear expert for one observable encoder state, not a surrogate for the official nonlinear detection logit. A universal local detector should be designed as a versioned union of independently calibrated experts, with family-wise false-positive control and an explicit abstention region. New experts still need source-matched clean counterfactuals or provider-oracle labels; metadata absence is not a negative label.

A symmetric transform-stability experiment challenged a simpler rescue. The weak oracle-positive 07.png scored 0.1248 natively, 0.0288 after JPEG 95, -0.0023 after JPEG 85, 0.1179 after a 0.99x Lanczos round trip, and 0.0791 after an eight-pixel crop round trip. The oracle-negative Open Images chestnut scored 0.2124, 0.1322, 0.0482, 0.2274, and 0.0601 under the same views. Both signals degrade similarly under codec and crop operations. Multi-view stability does not distinguish the weak positive from this natural lattice confound and is rejected as a rescue gate.

Raw folded-tile amplitude was also tested because 07.png had a norm of 6.924, compared with 3.017 for the chestnut and 0.874 for the strongest tested resized Kodak confound. The full corpus rejects a general amplitude gate. Among 4,698 Google rows, the current score threshold accepts 3,928, but accepted norms run as low as 0.712. A post-hoc branch at score >= 0.121 and norm >= 5 would add 190 rows below the current threshold, yet an observed-geometry development control already reaches score 0.120911 at norm 15.339. The margin to 07.png is under 0.004 and there is no untouched validation set. Amplitude is useful as a support diagnostic, but neither a veto nor this low-score rescue is eligible for runtime.

The next defensible Google experiment is consequently a preregistered multi-expert cascade: preserve the current carrier unchanged; train candidate experts only on source-grouped paired or oracle-labeled data; calibrate every expert and their union on untouched natural, resize, codec, and provider controls; and return unknown when neither the watermarked nor unwatermarked hypothesis clears its calibrated boundary.

The newly published MarkNull work supplies an independent no-box removal baseline. It reports 100% attack success on 20 Imagen-3 images checked through Gemini's verification flow. The attack uses a clean Stable Diffusion proxy to invert each image, optimize a bounded latent against a noise-latent alignment loss plus LPIPS, MSE, and SSIM, and decode it again. Its milder SynthID setting still reports only 25.36 dB PSNR and 0.80 SSIM. It therefore strengthens the case for latent regeneration as a detector-independent fallback, but it does not meet this project's pixel-only or fidelity gates.

The associated public repository contributed 50 metadata-free 512-by-512 Gemini/Imagen source candidates, but no attacked SynthID pairs in the inspected Git tree. The paper reports only 20 oracle-evaluated sources and does not map those verdicts to filenames, so all 50 remain source-asserted candidates rather than individual oracle labels. Both shipped experts detected zero of 50. More precisely, the fixed runtime expert reported unsupported because 512 by 512 is below its calibrated pixel-count floor. Its raw diagnostic score ranged from -0.1044 to 0.1058. Scale registration supported all 50 and accepted none, with its highest decision statistic at 0.9574 below the frozen 1.0 threshold. The shared scorer in scripts/synthid_runtime_expert_scores.py now records this distinction and exports both observations from one pixel decode.

An external-source split nevertheless exposed a highly reproducible period-4 component. Templates from the first and second 25 images correlated at 0.958. A template fitted only on the first half accepted 22 of 25 held-out candidates at a threshold above 250 canonicalized COCO development controls and accepted zero of a separate 250 COCO controls. The broader challenge rejected watermark specificity: it accepted two of 300 heterogeneous AI controls and five of 282 content-deduplicated internal negative-test controls at the same frozen threshold.

The mechanism diagnostics also identify the confound. The period-4 mean placed 82.4% of its non-DC spectral power in the horizontal Nyquist bin, and stable correlations at larger even periods vanished after projecting out all period-4 harmonics. Its strongest channel was Blue rather than Green, and its best phase-aware correlation with a fourfold-reduced shipped carrier was only 0.375. This is an axial renderer or resampling signature, not evidence for another validated diagonal SynthID carrier. It is retained as a hard-negative feature and excluded from the runtime expert bank.

A schema-versioned conformal cascade now implements the required aggregation in scripts/synthid_conformal_cascade.py. Every configured expert must emit an explicit supported score or unsupported result. For presence, the smallest empirical clean-null upper-tail p-value receives a Bonferroni correction over the supported expert bank. For absence, the largest lower-tail p-value against each expert's positive calibration is used, so every possible encoder state must be rejected. not_detected is disabled unless the calibration manifest explicitly declares complete version coverage and every expert supports the query. Input and output manifests are hash-pinned for reproducibility.

The first source-disjoint smoke test used only the unchanged fixed 16-by-16 expert. Alternating content hashes supplied 2,347 source-asserted Google calibration rows and 2,346 evaluation rows. The existing observed-geometry COCO partitions supplied 3,000 development controls and 2,000 final controls. At a two-sided alpha of 0.001 with incomplete-version coverage, the cascade detected 2,084 of 2,346 Google rows and one of 2,000 final natural controls; all other rows abstained and none returned not_detected. The corresponding frozen runtime threshold detected 1,954 Google rows and zero final controls. The conformal point therefore recovers 130 additional source-asserted positives at the cost of one natural false positive. It is statistically compatible with the target 0.1% FPR, but it is not a runtime replacement: the corpus has already influenced detector development, Google labels are provenance assertions rather than per-file oracle verdicts, and a multi-expert union still needs a fresh family-wise challenge.

The fixed and scale-registered experts were then evaluated together on their frozen source-disjoint resize challenge. Current production scores were recomputed from pixels for all 355 transformed Google candidates and 499 untouched Open Images controls. Sorting each label by source identifier and alternating rows produced 178 positive and 250 negative registered-calibration examples, followed by 177 positive and 249 negative evaluation examples. The fixed calibration used 4,516 Google sources after excluding every evaluation source and 2,980 observed-geometry development controls. At the unchanged runtime thresholds, fixed accepted 8 evaluation positives, registered accepted 116, and their union accepted 117 of 177 with zero of 249 controls. On the full descriptive corpus the corresponding counts were 15, 229, and 231 of 355, again with zero of 499 controls. Registered scale search therefore supplies 109 source-disjoint evaluation positives that the fixed expert misses; the fixed expert adds one that registered misses.

Applying the empirical cascade at alpha 0.01 was less specific. It accepted 122 of 177 evaluation positives but also five of 249 controls. All five control hits came from the registered branch below its frozen decision threshold. The 250-control calibration provides only 0.00398 p-value granularity, and lowering the operating point merely traded five additional positives for five false positives. This conformal operating point is rejected. The implementation is retained as versioned calibration infrastructure, while the independently frozen production thresholds remain the valid two-expert decisions.

A wider calibration reused the already saved component reports for 11,273 supported, symmetrically resized Spaces controls. The stored high-band value made the current registered decision score exactly reconstructible throughout the relevant upper tail; rows screened below the historical 0.1182 amplitude ratio cannot affect this tail. At alpha 0.001, the two-expert cascade accepted 119 of 177 evaluation positives and zero of 249 controls. The two additions over the runtime union had registered scores 0.9830 and 0.9731, with corrected clean-null p-values of 0.000710. This initially appeared to be a non-failing multi-expert calibration point, but the positive challenge had already been exposed. The alpha, expert bank, and thresholds were frozen before the fresh test below.

The decisive reserve used the next 1,500 sorted Open Images validation IDs, 3,500 through 4,999, frozen before scoring. Download and geometry checks yielded 1,000 registered-supported controls with no decoded-pixel hash overlap against the prior 1,499 Open Images controls. Registered scale search accepted zero. The fixed runtime threshold accepted 7 of the 81 controls in its supported geometry, or 7 of 1,000 overall, so a simple OR of the two experts also accepted seven. The frozen alpha-0.001 cascade was worse: it accepted nine controls, eight through the fixed tail and one registered-only row at score 0.8690. The candidate is rejected. In overlapping geometry, fixed-only evidence cannot be added to registered evidence without a new specificity mechanism; keeping the fixed detector available as a separately scoped diagnostic does not make their union safe.

The fixed-only geometry does not provide a fallback. Among 89 saved non-Google Spaces controls from 10 to 18 megapixels, where registered mode is unsupported but fixed mode is supported, 10 crossed the fixed threshold. The errors span multiple generator families. A geometry router therefore cannot extend recall by selecting fixed whenever registered is unavailable.

scripts/synthid_routed_expert_bank.py implements the surviving conservative policy without changing the standalone detectors. When registered mode is supported, only its frozen threshold can produce detected; a fixed-only crossing is retained in the report as fixed_only_ambiguous. When registered mode is unsupported, fixed evidence is likewise reported but the bank abstains. The bank never returns absence because encoder-version coverage is incomplete. On all 355 transformed positives and 499 reserve controls it detected 229 and zero. On the fresh 1,000-image reserve it detected zero and explicitly routed the seven fixed crossings to abstention. This is a safe expert-bank baseline, not a universal SynthID decoder: scale 0.5, unmeasured encoder states, and the separate OpenAI family still require additional validated experts.

Directly enlarging registered mode's pixel ceiling also failed. The unchanged registered components were computed on every saved 10-to-18-megapixel example: 37 Google candidates and 89 non-Google Spaces controls. The frozen threshold accepted one positive and zero controls. Amplitude passed for 28 positives and the selected spatial and spectral periods agreed for 21, but only three cleared the high-band gate. Large full-frame spectra therefore lose almost all recall; the absence of control errors does not justify widening the declared geometry. Phase-aligned window aggregation inside the calibrated range is the next bounded large-image hypothesis. That hypothesis was then challenged with every phase-aligned 2,048-square window. Median-window scoring retained 36 of 37 positives but also 10 of 89 controls; requiring every window to pass retained 36 positives and eight controls. The false lattices are spatially coherent, so window consensus cannot validate a large-image fixed fallback. Large images outside registered scope remain abstentions.

A final frozen full-frame alternative isotropically reduced each large image to eight megapixels with area interpolation before applying the unchanged registered detector. This maps a native period 16 into approximately 10.7-14.3 pixels, inside the calibrated period range, without discarding spatial regions. It still accepted only 1 of 37 positives and zero of 89 controls, and the one hit was the same 4096-square image detected natively. Amplitude passed for 26 positives, period agreement for 16, and high-band agreement for only three. Pre-resizing therefore does not recover the missing evidence and adds harmonic ambiguity; no target-size sweep is justified on this exposed cohort.

The scale-0.5 failure was also rechecked without the registered detector's sequential amplitude gate. High-band correlation was computed for all 65 period-8 positives and all 499 reserve controls. Positive and control medians were 0.0184 and 0.0213, with ranges -0.2381 to 0.1930 and -0.1734 to 0.2148. Normalized amplitude also overlapped: positive median and maximum were 0.4611 and 0.9259, compared with 0.3824 and 1.6296 for controls. Requiring amplitude and high-band evidence jointly therefore cannot recover this scale at the frozen specificity. The 0.5x case remains an explicit blind spot rather than an eligible lower-threshold branch.

2026-08-10: OpenAI periodic-carrier challenge

The OpenAI track repeated the Google carrier method without runtime provenance features. C2PA selected the research labels, but every candidate score consumed only decoded RGB pixels. The temporally split corpus contained 4,615 OpenAI training positives, 1,280 validation positives, 1,364 development-test positives, 8,668 difficult external-source controls, and 3,000 COCO controls. The development test is not called a locked holdout because earlier representations had already exposed it.

High-pass residuals were folded modulo 8x8, 16x16, and 32x32 and normalized as in the Google fixed-tile expert. A single 16x16 OpenAI template accepted only one of 1,280 validation positives at the threshold above all 2,287 validation controls. A 16-template geometry bank improved that to 12 positives but accepted one of 3,372 development-test controls. The corresponding development-test sensitivity was 9 of 1,364. The frozen Google template did not transfer: at its unchanged threshold it accepted 14 validation positives and 572 validation controls.

The apparent native-size signal was a confound. At 1024x1536, 91.6% of the learned template energy projected onto a 2x2 periodic lattice, compared with 5.6% for the frozen Google template. Firefly, China AIGC-labeled, local diffusion, and unidentified hosted-generator controls reached correlations of 0.92-0.98 against the uncorrected OpenAI template. This is operational evidence for a shared decoder or upscaler lattice, not an OpenAI watermark carrier.

Projecting the 2x2 nuisance lattice out of every 16x16 fold isolated a weaker epoch-specific residual. The 1024x1536 expert accepted 5 of 55 validation and 10 of 51 development-test positives while accepting none of the full validation or development-test controls. Nine of the ten development-test matches occurred from July 12 through July 14; later dates contributed only one of 27. The 1536x1024 expert accepted 6 of 37 validation and 2 of 51 development-test positives with no development-test controls. Amplitude, active-support, channel-consensus, payload-invariant power, and spherical multi-template branches did not improve this frontier without false positives.

The OpenAI branch therefore remains research-only and is not added to the runtime detector. The next valid training epoch requires same-content OpenAI-oracle detected and not_detected pairs, or provider-authorized decoder labels, followed by a fresh temporal holdout. Source provenance alone cannot distinguish a true watermark carrier from the measured generator lattice.

2026-08-12: OpenAI opponent-color carrier challenge

The RGB periodic and learned branches did not test the hypothesis that a weak carrier is clearer after the nonlinear LMS transform used by OKLab. A new validation-only screen compared sRGB, linear RGB, cube-root LMS, OKLab, and an orthonormal linear opponent basis. For each representation it tested periods 4, 8, and 16 plus whole-vector, single-channel, and chroma-vector folds. Every candidate fitted a positive-minus-control direction independently in the native and JPEG-95 views. A positive had to cross both thresholds, with each threshold strictly above every validation control in the corresponding view.

The development corpus was limited to the three OpenAI native geometries: 1024x1024, 1024x1536, and 1536x1024. It contained 283/140/158 OpenAI train/validation/test images and 252/145/135 heterogeneous controls. The best validation candidate was the third cube-root LMS component at period 8. It accepted 19 of 140 validation positives. The frozen locked test accepted 16 of 158 positives and none of 135 controls in the conjunctive native-plus-JPEG rule. Native alone accepted 43 positives and one control; JPEG-95 alone accepted 22 positives and a different control. The native and JPEG templates correlated at 0.890.

The 16 test hits were not a single date, geometry, or content family. They spanned 11 collection dates; the 1024x1024, 1024x1536, and 1536x1024 cohorts contributed 2, 8, and 6 hits. A 256-bit DCT-hash audit found two multi-member perceptual groups in the complete 581-image OpenAI native cohort, but neither connected an accepted test image to another split. Thresholding against the union of heterogeneous validation controls and 140 matched 0.75x-down-and-up OpenAI views did not change the operating point. The 158 matched test views accepted none while their 16 corresponding source views remained accepted.

The transform boundary is mixed. All 16 source hits remained conjunctively accepted after WebP-95; the complete test set then contained 26 accepted OpenAI images and one accepted control. JPEG-90 retained four of the 16. Crop and 0.75x resize round trips retained none. Searching all 64 cyclic period-8 phases did not restore geometric robustness: resize retained 2 of 158 positives and one of 135 controls, while crop retained no positive and one control. The loss is not merely an unknown modulo origin.

The candidate was also scored without refitting on all 52 action-derived same-provider controls at these geometries. It accepted one of 52 and none of the 12 controls explicitly naming gpt-image 2.0. The exact-version subset is only 8/2/2 train/validation/test by the native geometry restriction, and the absence of c2pa.watermarked.unbound remains an action-derived candidate label, not a provider-oracle negative. These counts cannot establish an OpenAI-specific false-positive rate.

The dominant exporter lattice was removed explicitly. The raw period-8 template placed 84-90% of its energy in period-2 or period-4 subspaces. For each image, projecting out the complete period-4 component before refitting reduced native/JPEG template correlation to 0.564 but still accepted 18 of 140 validation and 15 of 158 test positives with zero test controls. The useful residual therefore does not collapse to the previously documented 2x2/4x4 generator lattice.

A direct pixel ablation tested causality against the frozen local score. The normalized native period-8 direction was subtracted only from the sRGB Blue channel, with a scalar search for the smallest amplitude that cleared both the native and JPEG-95 thresholds. All 16 locked-test sources cleared. Repeating the same selected amplitude at the least-correlated cyclic phase cleared none of 16. Median amplitude was two Blue-channel integer levels, median PSNR was 52.24 dB, and median SSIM was 0.99639. The minima were 40.62 dB and 0.97184, so even as a local ablation this is not yet a production-quality remover. More importantly, score causality does not establish that the candidate is the OpenAI SynthID decoder surface.

An all-size map then evaluated the same fixed native period through partial modulo folds, so dimensions did not need to be divisible by eight. It accepted 94 of all 7,259 OpenAI-labeled files across 13 of 605 geometries and 9 of 8,670 heterogeneous controls across 7 of 1,534 geometries. The three fit geometries contributed 76 positives. The largest exported geometries, including 2,265 files at 1086x1448 and 992 at 1254x1254, contributed none. Two corpus rows were excluded because their decoded geometry no longer matched the indexed metadata.

A scale search was frozen to integer periods 4 through 16 before examining the all-size results. Its score was the maximum, across periods, of the minimum native and JPEG-95 correlations. The common threshold was set above every period-searched validation control, thereby charging the search itself against specificity. It accepted 13 of 1,854 validation positives and none of 2,227 validation controls. On test it accepted 14 of 1,839 positives and one of 2,162 controls. Every accepted OpenAI image selected period 8; no other period added a hit. The carrier does not scale continuously with stored resolution.

This is the strongest local OpenAI periodic candidate so far: it is codec tolerant, survives explicit subperiod deflation, has a phase-specific causal ablation, and is not explained by date, native geometry, or hash-group leakage. It is still research-only. It has low sensitivity, fails geometric transforms, has one all-size test-control collision, and lacks a matching-provider oracle. The next independent branch should look for a resampling-invariant, content-dependent residual on source-grouped data rather than broadening the same period search. No runtime OpenAI pixel expert or remover is justified.

The next branch removed phase from the folded LMS spectrum and required three views: native pixels, JPEG 95, and a 0.75x down-and-up resize. It screened magnitude, power, and log-magnitude profiles at periods 4 through 16 on validation only. The selected period-16 power candidate used a separately fitted template for each view. It accepted 8 of 140 validation positives and 8 of 158 locked-test positives, with no native-geometry control accepted in either split. On the complete wide challenge it accepted none of 2,227 validation and 2,162 test Spaces controls, one of all 261 action-derived same-provider controls, and none of the 118 exact-version gpt-image 2.0 controls. Five of its eight test hits overlapped the signed period-8 candidate; three were additional.

Subperiod deflation bounded that result. Removing the exact period-8 Fourier support from each period-16 spectrum improved native-to-resize template correlation from 0.061 to 0.925, but the locked test then retained five positives and accepted two of 135 controls. Removing period-4 support retained no test positive. The transform-tolerant energy is therefore part of the same harmonic family rather than a clearly independent carrier.

Finally, a stricter experiment required one shared phase-free template across all three views. Validation selected period-15 magnitude, whose independently measured view directions correlated 0.992 for native versus JPEG and 0.900 for native versus resize. It retained 7 of 158 native-geometry locked-test positives and none of 135 controls. That apparent invariance failed on the wide geometry challenge: the same thresholds accepted 11 of 2,227 validation and 14 of 2,162 test Spaces controls, plus seven of 261 same-provider controls and two of 118 exact-version controls. It is a general renderer or resampling spectrum and is rejected.

The separately fitted period-16 three-view ensemble remains a useful research diagnostic because it survived the broader control challenge, but it is not a single codebook, has only eight native-geometry test positives, and derives most of its signal from the same subperiod family. It does not rescue the universal-detector claim or justify runtime code.

A non-learned fractal hypothesis was tested last. For each image it compared the phase-free cube-root LMS spectrum with the same image after JPEG 95 and after a physical 0.75x reduction. Candidate native periods 8, 12, 16, 20, and 24 were paired with scaled periods 6, 9, 12, 15, and 18. The score was the minimum of native-to-JPEG and native-to-scaled spectral cosine, so no provider template or fixed phase participated. Validation selected period 12-to-9 log-magnitude and retained 3 of 140 positives above all 145 controls. The locked test accepted none of 158 OpenAI positives and one of 135 controls; the OpenAI median was 0.515 versus 0.555 for controls. Simple phase-free spectral self-similarity is not an OpenAI watermark feature and the fractal branch is rejected.

The scale hypothesis was then challenged against the common non-native export geometries. A nearest-native inverse-raster surrogate recovered none of 1,321 test OpenAI images and accepted two of 1,485 symmetrically resampled control views. That mapping was intentionally treated as incomplete because it changed aspect ratio for important cohorts. The largest cohort, 1086x1448, is instead consistent with an isotropic 1/sqrt(2) reduction from 1536x2048. Restoring that exact hypothesized source raster still recovered none of 579 validation or 562 test positives. A direct fractional-phase fold avoided inverse interpolation and searched two raster-coordinate conventions plus all 64 cyclic origins, charging the complete registration search against validation controls. It retained 2 of 579 validation and 4 of 562 test positives while accepting 2 of 135 test controls. Its median score was lower for OpenAI than for controls. The fractional-carrier explanation for this dominant export geometry is rejected.

A nonlinear cyclostationary screen next tested whether content-dependent or payload-varying signs cancel the cross-image mean while leaving a periodic variance. It compared folded magnitude, energy, locally normalized energy, and cross-channel products in cube-root LMS and OKLab. Validation selected the absolute residual of the third cube-root LMS component at period 4. It retained 33 of 140 validation and 43 of 158 locked-test positives, but also accepted 2 of 135 locked-test controls. The frozen wide challenge confirmed a renderer confound: it accepted 48 of 1,854 validation and 52 of 1,839 test OpenAI images, against 7 of 2,227 validation and 5 of 2,162 test Spaces controls. It also accepted 10 of 261 action-derived same-provider controls, including 5 of 118 that explicitly name gpt-image 2.0. Requiring spatial agreement did not repair specificity; the validation-selected median over four quadrants kept 41 of 158 test positives and the same two native-geometry test controls.

Finally, a second-order cyclic screen folded spatial products of the residual with lags 1, 2, 4, and 8 along axial and diagonal directions. It evaluated 360 fixed LMS/OKLab, channel, period, lag, and normalization combinations. The validation winner used the second cube-root LMS component, period 16, and a vertical four-pixel lag. It retained 36 of 140 validation and 43 of 158 test positives, but accepted 4 of 135 locked-test controls. Periodic magnitude, patch consistency, and lag autocorrelation therefore expose reproducible OpenAI renderer structure but do not meet the SynthID specificity gate.

These negative results bound the current manual feature search. Signed phase, phase-free power, subperiods, integer and fractional scale, inverse raster registration, spectral self-similarity, nonlinear periodic variance, spatial consistency, and second-order lag products have all either lost sensitivity or crossed held-out controls. Adding further hand-selected transforms to the same exposed corpus would increase adaptive overfitting rather than confidence. A defensible next OpenAI epoch requires provider-authorized oracle labels or clean same-content counterfactual pairs and a fresh temporal holdout.

An untouched 2026-08-07 through 2026-08-10 temporal cohort supplied that holdout for validation, but not the missing oracle negatives. It contained 486 explicit OpenAI-provenance positives and 504 heterogeneous controls. Only seven additional files were pure OpenAI C2PA candidates without a SynthID assertion; the remaining same-platform candidates contained inferred or mixed Google and OpenAI watermark provenance and were not treated as clean negatives. The frozen signed period-8 candidate accepted 5 of 486 positives and 1 of 504 controls. The phase-free period-16 candidate accepted 2 of 40 supported positives and 1 of 50 supported controls. The nonlinear period-4 candidate accepted 14 of 486 positives and 8 of 504 controls. The manual carrier branches fail temporal transfer.

2026-08-10: OpenAI content-dependent decoder challenge

The periodic result did not exhaust the neural-watermark hypothesis. OpenAI's ChatGPT Images 2.0 system card describes its watermark as content-specific, and the SynthID-Image paper says the encoder is content-dependent by design. The paper also separates watermark detection from payload recovery and evaluates the external SynthID-O model at its preferred 512x512 input size. These facts support a learned full-field decoder, but they do not establish that OpenAI and Google deploy identical model weights, payloads, or pixel spectra.

A second corpus audit found 261 same-provider candidate controls. Their signed provenance contained created and converted actions without the c2pa.watermarked.unbound action, but a subsequent software-agent audit showed that they do not all come from the positive generation epoch. Only 118 name gpt-image 2.0: 97 training, 16 validation, and five development-test images. Another 108 name an earlier version and 35 do not expose a usable version. The full 261-image split remains 217/26/18. These are useful confound controls, but even the exact-version subset is action-derived rather than a provider-oracle negative label.

Two pixel-only native-resolution experts were trained with the mixed-version control set. A local residual CNN accepted 132 of 1,280 validation positives and 149 of 1,364 development-test positives, with zero same-provider or COCO controls in both splits and two of 1,373 difficult development-test controls. A separate spectral, autocorrelation, bit-plane, and residual-error expert accepted 175 validation and 167 development-test positives with no accepted controls. A validation-selected linear combination accepted 243 positives in each split, with no accepted validation or development-test control. The zero-error development-test bound across 3,391 controls is 0.0883% at one-sided 95% confidence, but the same-provider subset contains only 18 images and therefore does not independently establish provider specificity.

The exact-version audit did not invalidate the native score counts, but it weakened their attribution. None of the 16 validation or five development-test gpt-image 2.0 controls crossed the residual, forensic, ensemble, canonical, or JPEG-equalized thresholds. That is directionally useful, but five temporal test controls cannot establish deployment specificity. Perceptual-hash and local CLIP searches compared all 118 exact-version controls with 7,259 positives. They recovered a few near-content matches but no clean before/after watermark pair. Sub-pixel affine alignment left content, edge, and photometric changes in the residual, while 8x8, 16x16, and 32x32 folded residual correlations varied in sign and magnitude. Those candidates cannot be used as causal watermark differences.

Transformation challenges rejected the combined native expert as a SynthID decoder. Of its 243 accepted development-test positives, 121 remained accepted after a 5% crop round trip, six after a 75% resize round trip, and none after a JPEG-95 round trip. No transformed same-provider control was accepted. The codec fragility is inconsistent with treating the score as the documented robust watermark without further evidence.

The 512x512 branch also failed the held-out specificity gate. Its selected model accepted 238 validation and 194 development-test positives, but also accepted two of 18 same-provider, four of 1,373 difficult, and nine of 2,000 COCO development-test controls. Finally, a CNN trained and evaluated only after applying the same JPEG-95 round trip to every positive and control accepted eight validation positives and six development-test positives while accepting five of 2,000 development-test COCO controls. Equalizing the codec therefore removed nearly all of the apparent signal instead of revealing a robust hidden decoder feature.

An independent three-CNN OpenAI surrogate was also evaluated from its published weights, without adopting its code or claims. At the published 0.5 threshold its ensemble marked 82.0% of COCO validation controls and 83.4% of COCO development-test controls as watermarked. It also marked 56.3% of difficult development-test controls and 15 of 16 exact-version OpenAI validation controls, compared with 87.3% of OpenAI validation positives. The repository does not publish its training corpus or logs, and its implementation applies supposedly shared pair augmentation independently. This result rejects the model as a SynthID decoder and demonstrates why same-content-looking regenerations are not sufficient negative labels.

A final transform-augmented full-field branch tested the closest practical approximation to the documented neural decoder. A 512x512 CNN consumed both RGB and a fixed high-pass residual and trained with random crop/resize, JPEG 70-100, horizontal flip, and mild color transforms. Training sampling balanced watermarked positives, difficult controls, and the exact-version OpenAI candidate controls. Selection required a score above every control for both native and JPEG-95 validation views. All four fixed training epochs accepted zero of 500 selection positives. On the full validation set the frozen model accepted zero of 1,280 OpenAI positives in native, JPEG-95, 75%-resize, and 5%-crop views, while accepting up to two unrelated controls. On development test it accepted zero of 1,364 OpenAI positives in native, JPEG-95, and resize views. The crop view accepted one OpenAI positive and one of 1,373 difficult controls. Transform augmentation therefore did not recover a usable weaker OpenAI signal.

The measured OpenAI-native separation is best classified as an export or rendering noiseprint, not a validated SynthID detector. None of these models is added to runtime. Transform-augmented training has now also failed. A defensible next epoch needs provider-authorized labels or clean same-content before/after watermark pairs, plus a fresh temporal holdout that has not influenced representation or threshold selection.

The same untouched August cohort then challenged the frozen 512-square forensic model and its unchanged threshold. It accepted 82 of 486 explicit OpenAI positives, one of 504 heterogeneous controls, and one of seven pure OpenAI C2PA candidate controls. The aggregate score separation was real, but transformation testing resolved its interpretation: all 82 native detections disappeared after both JPEG-95 and WebP-95 round trips. A 0.75x resize retained 46, and a 5% crop retained 23. The model therefore generalizes as a fragile OpenAI export noiseprint, not as the documented robust SynthID signal.

The provider-key alternative was then tested directly. OpenAI training folds were clustered into 1, 2, 4, 8, 16, or 32 spherical period-8 cube-root-LMS directions, and the cluster count was selected on validation only. Eight directions raised the development-test result from 16 to 27 of 158, with one of 135 heterogeneous controls and two of 52 action-derived same-provider controls accepted. This did not reveal eight independent codewords: the two dominant clusters contained 234 of 283 training rows, and their templates correlated at 0.982 natively and 0.924 after JPEG-95.

The frozen multi-direction model accepted none of 443 Google native-geometry files, but accepted 5 of 39 Microsoft Designer files and 123 of 674 OpenAI- platform files. All five Microsoft hits had joint Microsoft, OpenAI provenance, one explicitly asserted an OpenAI SynthID watermark, and all five selected the same dominant direction as 111 OpenAI hits. On the fresh August native-geometry cohort, SHA-256 grouping reduced seven OpenAI row hits to five unique content groups; two unique Microsoft controls crossed the same dominant direction. A sign-invariant carrier-subspace fit retained only 1 of 158 development-test positives and none of 40 fresh positives. A final OpenAI direction orthogonal to Google, Microsoft, and other-provider train means likewise retained 1 of 158 and none of 40. The measured period-8 family is distinct from Google but is shared with a Microsoft/OpenAI lineage.

Regrouping by the signed watermark assertion found 115 of 589 explicit OpenAI- SynthID files accepted, against 13 of 117 OpenAI-lineage files without an assertion, none of 443 Google files, and one of 89 other-provider files. The aggregate one-sided exact-test result was p = 0.0179; the already exposed development-test split was not independently significant at 27 of 159 versus 2 of 25 (p = 0.202). Missing watermark action is not an oracle-negative label, so this is compatible with a provider-keyed family but does not prove that the measured component is the watermark.

The frozen score did exhibit real codec persistence. WebP-95 retained 24 of 27 native development-test detections and added nine other positives, with the same one of 135 controls retained. JPEG-90 retained eight of 27 and added no positive, with zero controls. A 0.75x resize and 5% crop retained none. The result is stronger than the full-image noiseprint and supports a genuine OpenAI/Microsoft raster-phase component, but its geometric fragility excludes it as the documented complete robust SynthID decoder surface.

The broader SynthID architecture permits provider-specific codes without requiring one shared pixel template. The SynthID-Image paper separates binary detection from payload recovery, identifies payloads as a mechanism for distinguishing customers of one service, and assigns watermark versions to the encoder. OpenAI and Google can therefore use the same watermark family while deploying different encoder versions or payload distributions.

The narrower same-carrier/different-phase prediction was challenged directly. Period-8 cube-root LMS folded residuals were collected from 581 byte-unique OpenAI-asserted files, 443 Google-asserted files, 93 OpenAI-lineage files with no watermark assertion, and 82 other-provider controls. Both provider contrasts were internally stable across deterministic halves. Raw phase-free power means had cross-provider cosine 0.480, but this shared imaging background disappeared after subtracting the respective controls: positive power support had cosine 0.0123, bootstrap median 0.0232 with a 95% interval of 0.0009 to 0.0856, and no overlap among the 12 strongest coordinates. The three per-channel cosines were all at most 0.0155. The signed contrasts were negative, not phase-locked. The evidence therefore rejects a simple Google template with an OpenAI phase or payload substitution. It remains compatible with a distinct, content-dependent SynthID encoder version and motivates separate learned experts rather than further rotation of the Google carrier.

A fixed three-query OpenAI Verify validation then established attribution while also falsifying completeness. An asserted OpenAI image accepted by the frozen eight-direction period-8 candidate was submitted both in its original form and as a pixel-identical PNG re-encoded without metadata. Both returned SynthID detected and Content Credentials not detected. A second asserted OpenAI image rejected by the local candidate, with native/JPEG scores 0.709/0.657 against thresholds 0.918/0.923, was also re-encoded without metadata; it returned the same pixel-only SynthID verdict. The period-8 family is therefore a genuine OpenAI SynthID component with known oracle-positive false negatives, not merely an export fingerprint and not a complete detector. No adaptive oracle sweep followed: OpenAI's Content Provenance documentation explicitly disallows repeated queries to reverse-engineer, remove, or evade a watermark. The next local epoch must treat period-8 as one expert and discover complementary content-dependent experts using frozen corpus splits, without oracle-guided feature selection.

Relaxing the eight-direction model to a separate validation-control threshold for every direction increased validation recall to 36 of 140 with zero of 145 controls, but it did not preserve specificity. The locked development test accepted 46 of 158 OpenAI rows and 4 of 135 controls. A wider provider challenge accepted 180 of 603 explicit OpenAI-SynthID rows, 15 of 80 OpenAI-lineage rows without an assertion, 6 of 443 Google rows, and 7 of 112 other-provider rows. The fresh temporal result was 13 of 40 OpenAI rows, 4 of 50 general controls, and 1 of 3 same-provider controls. The oracle-positive local false negative did cross direction 4 under this relaxed calibration, but that direction survived only 1 of 4 leave-one-positive-out refits; the oracle-positive row itself then failed its JPEG threshold. Direction-specific thresholds therefore expose unstable minority clusters rather than additional validated codewords.

A train-only bilinear follow-up tested whether a content-dependent or keyed phase varies between images while retaining native-to-JPEG covariance within each image. It subtracted the control within-class cross-covariance from the OpenAI cross-covariance, selected the positive-eigenvalue rank on validation, and kept the threshold above every validation control. Rank 1 retained only 1 of 158 locked-test OpenAI rows and zero of 135 controls, then zero of 40 fresh OpenAI rows and zero of 53 fresh controls. A single codec-stable variable-phase subspace is therefore rejected as the missing complementary expert.

Taking local spectral magnitude before spatial averaging produced a superficially stronger complementary signal. Eighteen period, block-normalization, and spectral variants were screened on validation; period-8 unit-block log-magnitude retained 8 of 158 locked-test OpenAI rows and zero of 135 controls, then 3 of 40 fresh OpenAI rows and zero of 53 fresh controls. None of the three fresh hits overlapped the frozen global eight-direction hits, and only one overlapped the unsafe per-direction union. Both oracle-positive images, including the known local false negative, remained below this expert's thresholds.

The full provider challenge rejected watermark attribution despite that clean development transfer. The local-spectrum expert accepted 17 of 603 explicit OpenAI-SynthID rows, 5 of 80 OpenAI-lineage rows without an assertion, zero of 443 Google rows, and 2 of 112 other-provider rows; both other-provider hits were Microsoft. Explicit watermark assertion was not enriched over same-lineage non-assertion (p = 0.966, one-sided exact test in the required direction). Recalibrating both frozen view thresholds above the 32 same-lineage validation maxima reduced locked-test recall from 8 of 158 to zero. Local pre-aggregation magnitude therefore reveals another OpenAI/Microsoft raster component, not a defensible complementary SynthID expert.

The proposed directional-wavelet branch was also tested directly on OpenAI. A three-level DTCWT over the period-8 cube-root-LMS residual screened all six orientations, the diagonal and axis-near groups, individual levels, and complex, magnitude, and phase statistics. Validation selected all-orientation normalized complex correlation at 19 of 140 OpenAI rows and zero of 145 controls. It retained 16 of 158 locked-test OpenAI rows and zero of 135 controls, followed by 2 of 40 fresh OpenAI rows and zero of 53 fresh controls.

This is a stronger representation of the known phase carrier, not a new expert. Both fresh hits were already accepted by the frozen eight-direction model, the known oracle-positive local false negative remained negative, and the other oracle-positive row missed only the DTCWT JPEG threshold at 0.894 versus 0.897. The provider challenge accepted 77 of 603 explicit OpenAI-SynthID rows, 3 of 80 OpenAI-lineage non-assertions, zero of 443 Google rows, and 2 of 112 other rows, both Microsoft. The explicit-assertion enrichment was real (p = 0.00879), but 72 of the 77 hits overlapped the signed phase expert and their union increased its explicit count only from 115 to 120. Same-lineage calibration retained 4 of 158 locked-test and 1 of 40 fresh OpenAI rows, while one Microsoft test row still crossed.

The transform challenge then falsified operational shift invariance. Cyclic rolls by (1, 1) and (3, 5), a 0.75x resize round trip, and a 5% crop round trip each retained zero of the 16 baseline detections. JPEG-90 retained four. WebP-95 retained 15, accepted 25 OpenAI rows in total, and introduced one of 135 controls. DTCWT therefore improves codec tolerance for the already measured raster phase but does not solve registration, scale, arbitrary resolution, or the missing OpenAI payload variants.

Selecting the best DTCWT magnitude and phase candidates within their own families did not change that conclusion. The off-diagonal magnitude candidate retained 10 of 158 locked-test OpenAI rows and zero of 135 controls, then 3 of 40 fresh OpenAI rows and zero of 53 controls. All three fresh hits were already signed-phase hits. The near-vertical phase candidate retained 9 of 158 and zero of 135, but failed temporally at 1 of 40 OpenAI rows versus 3 of 50 general controls. Magnitude also retained zero of its ten baseline detections after a one-pixel roll, (3, 5) roll, resize, or crop; JPEG-90 retained one and WebP-95 retained two. Taking magnitude only after periodic complex folding preserves the phase-origin dependency and does not realize DTCWT's intended shift invariance.

A final DTCWT screen removed periodic folding entirely and summarized global energy over six orientations and three levels. Nine validation-screened translation-insensitive magnitude, power, orientation, and robust-quantile statistics selected median magnitude proportions within each level. The locked test accepted only 5 of 158 OpenAI rows and 1 of 135 controls; the fresh holdout accepted zero of 40 OpenAI rows and zero of 53 controls. DTCWT therefore presents a measured tradeoff in this corpus: preserving the periodic complex field retains the real OpenAI carrier but its phase origin, while pooling enough to be translation insensitive removes the transferable separation.

A metadata-stratified audit then tested whether the 77 provider-challenge hits were confined to one encoder rollout or geometry. Within the exact OpenAI Media Service API cohort, the frozen DTCWT expert accepted 75 of 581 asserted rows and 1 of 52 same-generator rows without an assertion. Asserted-row rates were 3/16 in May, 46/365 in June, and 26/200 in July, and 6/47, 44/320, and 25/214 across the three native geometries. The native/JPEG minimum score had 0.721 AUC against the same-generator non-assertion stratum. The component is therefore continuous across time and geometry rather than a short-lived export cohort, although missing assertion remains an indeterminate label rather than a clean watermark negative.

A prespecified shift-invariant follow-up sorted all 64 cyclic period-8 template correlations within each image. This preserves the shape of the correlation orbit while discarding its absolute phase; one train-only contrast was fitted for native and JPEG-95 views, with both thresholds placed above all 145 validation controls. The individual views retained 5 and 7 of 140 validation positives, but their conjunction retained none of 158 locked-test positives and none of 135 controls. JPEG, WebP, resize, and crop could not rescue a baseline detection because there was none. Thus the transferable OpenAI evidence is carried by signed raster phase, not merely by a shift-invariant matched-filter peak shape.

Content-hash grouping also corrected the temporal denominators. The 486 OpenAI rows represented 417 unique byte groups, and the 234 Google rows represented 211. The signed OpenAI period-8 result became 4 of 417 unique positives versus 1 of 454 general-control groups; the full-image noiseprint became 68 of 417; the Google runtime result became 153 of 211. Repeated uploads did not change the decision, but row-level counts must not be treated as independent evidence.

2026-08-13: OpenAI low-texture phase and masking-alignment challenge

The dominant 1086x1448 export geometry supplied 2,265 byte-unique images with an explicit OpenAI SynthID assertion, 830 general controls at the same geometry, and 26 OpenAI-lineage files without a watermark assertion. The exact-geometry split contained 1,124/579/562 asserted OpenAI train/validation/test rows and 399/203/228 general controls. The lineage rows remained an indeterminate stratum rather than being relabeled as negatives.

A train-only cross-image phase model selected 256 RGB bins from the 20 lowest-texture references in each of three luminance groups. It used polarity-invariant axial phase and scored both native pixels and a JPEG-95 view after canonicalizing to 512 square. Independent train halves shared only two of 256 exact channel-frequency bins and three spatial coordinates. The full model produced native AUC 0.526 on validation and 0.512 on test; JPEG-95 AUCs were 0.521 and 0.512. A threshold above every validation control accepted none of 562 test positives, while a validation 1% FPR threshold accepted four positives and two of 228 test controls. Texture quartiles did not reproduce a useful exception. The common absolute-phase hypothesis is rejected for this OpenAI geometry.

The official architecture points to a different mechanism: watermark generation is deliberately content-dependent to resist watermark exchange, and nearly uniform content can be filtered rather than watermarked. A second frozen probe therefore discarded absolute phase and global residual amplitude. It measured 60 OKLab relations between blockwise high-frequency energy and local texture, contrast, and chroma at two filter scales and two block sizes. One train-only shrinkage-Fisher direction was scored in native and JPEG-95 views, and the joint decision used their minimum above every validation control. Joint AUC was 0.563 on validation and 0.593 on test. The frozen test count was 15 of 562 asserted OpenAI rows and one of 228 controls.

That weak enrichment was not a stable expert. Directions fitted on deterministic train halves had raw-coordinate cosine 0.148. One half accepted 13 test positives and two controls, the other accepted none; their test score correlations were 0.335 on positives and 0.127 on controls, with no overlapping positive decision. The branch is rejected as an unstable content or export correlate. Together these experiments reinforce the existing boundary: a new OpenAI epoch needs pixel-oracle labels or clean same-content counterfactuals, not further adaptive feature selection on provenance-labeled exports.

2026-08-13: multi-scheme domain generalization and Gemini bicoherence

A deliberately broader experiment tested whether SynthID could inherit a generic post-hoc watermark signature instead of a provider carrier. Ninety source-matched COCO pairs were prepared for each of TrustMark P, Meta VideoSeal Y-256, classical 64-bit DWT-DCT, and Meta WAM-MIT, always with random payloads. The official VideoSeal decoder recovered a mean 99.6% of bits across identity, JPEG-90, 0.8x resize round trip, and 5% crop round trip; its mean pair PSNR was 45.04 dB. WAM recovered 99.5-99.6% across the same views at mean PSNR 38.44 dB. DWT-DCT supplied an architecture-diverse control, but only identity and resize were retained as positives because its official decode collapsed to chance after JPEG and crop.

A sign-symmetric 3,900-feature spatial rich model first learned residual co-occurrences in luminance, opponent-color, and Lab channels. Its linear and quadratic paired-delta variants recognized familiar schemes but did not transfer to a held-out scheme. A stronger neural follow-up used a fixed residual bank, worst-family training, cross-family residual transplantation, random sign, phase, and amplitude, plus phase-scrambled equal-power residuals as hard negatives. Each of four runs withheld one complete encoder family and all test source images. Held-out aggregate AUCs were 0.562 for TrustMark, 0.474 for VideoSeal, 0.497 for DWT-DCT, and 0.437 for WAM. The corresponding strict true-positive rates were 5.0%, 3.3%, 0%, and 0%. This rejects the available open-proxy path to a universal single-image detector: even causal paired data from three distinct schemes did not generalize to a fourth.

A separate Gemini experiment tested a translation-invariant higher-order signature. Block bicoherence measured normalized phase coupling at period-16 triplets satisfying f1 + f2 = f3; unlike ordinary phase correlation, the global image-shift phase cancels from this product. Six Green, cube-root LMS, and OKLab representations, two residual scales, two block sizes, and three train-only harmonic counts produced 72 candidates. The data were separated into harmonic-fit, candidate-selection, threshold-calibration, and final splits: 30/20/50 Gemini positives and 100/100/100/199 Open Images controls. The best selection candidate accepted none of 20 positives at the selection control maximum and had AUC 0.523. After independent recalibration, the final result was 0 of 50 Gemini positives and 0 of 199 controls, with AUC 0.374. Period-16 bicoherence is therefore rejected as a scale or origin rescue. The measured Gemini carrier remains a linear, registered periodic tile rather than a transferable harmonic cascade.

2026-08-14: remaining Gemini resolution challenges

The 0.5x gap was tested with a local-consensus detector rather than another global period-8 score. Each half-scaled image was divided into 64-pixel patches; the 90th percentile of full-template patch correlations was selected without using the locked test. It accepted 33 of 49 validation positives and zero of 166 validation controls at the frozen 0.552176 threshold, then 27 of 52 locked-test positives and zero of 140 controls. This looked like a useful low-resolution expert but did not survive scale-matched specificity testing.

The frozen challenge symmetrically resized every remaining Spaces control and a separate 1,000-image Open Images reserve to 0.5x. The expert accepted 61 of 10,906 Spaces controls and 22 of 1,000 Open Images controls. All 83 errors were rescored with the prespecified feature inventory. A negative-Green patch-spread gate could remove them post hoc, but retained only 12 of 106 base-positive development hits. On the later August 7-10 cohort the original patch expert accepted 67 of 234 Google-provenance rows and 8 of 689 controls, while the new conjunction retained only 2 Google rows and zero controls. The broad challenge rejects the base expert and the temporal result rejects the post-hoc repair. No 0.5x claim is added.

The large-image window family was completed separately. One centered, phase-aligned 2,048-square registered window accepted 1 of 37 candidates above 10 megapixels and zero of 89 controls. Scoring every phase-aligned 2,048-square window and accepting the maximum raised recall to only 2 of 37, still with zero controls; the maximum control decision score was already 0.9677 against the 1.0 runtime boundary. Together with the previously rejected fixed-window consensus and full-frame resize, these results leave the 10-megapixel ceiling intact.

Near-ceiling windows, color spaces, and JPEG-paired phase directions did not provide a missing large-image expert. Three aligned windows of approximately 10 megapixels retained 1 of 37 candidates and zero of 89 controls. A native fixed-score plus Green mid-band rule retained 15 of 19 candidates in four exact recurring geometries and accepted one of 11,417 resized Spaces controls, but collapsed to 0 of 19 after JPEG 90. A codebook fitted to the minimum native/JPEG period-16 correlation selected RGB Green mid-band on the pre-July split. It retained 23 of 26 candidates and zero of 40 controls there, then 10 of 11 and 7 of 49 on July. Lab and OKLab candidates failed on the same hard control. The apparent native signal is a codec or upscaler lattice rather than a robust watermark invariant.

Those 37 candidate labels came from Google C2PA and an inference that Google media is likely watermarked, not an individual Google pixel-decoder result. They cannot support a production recall claim. The next large-image step is an immutable SHA-to-verdict Gemini batch, followed by fitting only on oracle-detected files. The first live attempt on 2026-08-14 was stopped by Google reCAPTCHA before the first new file submission; no CAPTCHA bypass was attempted.

2026-08-14: official large oracle and opponent-color runtime branch

The frozen Gemini batch was subsequently submitted without bypassing CAPTCHA. Every submitted file had AI provenance removed while preserving the decoded RGBA fingerprint. Google returned a pixel-positive verdict for seven of seven usable submissions across 3072x5504, 5504x3072, 5632x3072, 6336x2688, 3584x4800, 5016x3328, and 4096x4096. An eighth case was indeterminate after the verifier quota ended. This is direct evidence for seven positives, not a label upgrade for the remaining 30 C2PA-inferred candidates.

Revisiting the native period-16 fold by opponent color exposed the missing separation. Google candidates preserve strong Red-minus-Green and Blue-minus-Yellow phase agreement across every phase-aligned 2048-square window, while natural resize aliases do not preserve the complete conjunction. The frozen large-v1 rule requires minimum fixed score 0.14, minimum Red-minus-Green spatial correlation 0.90, minimum Blue-minus-Yellow spatial correlation 0.70, and a minimum signed Blue-minus-Yellow mid-band correlation at or below -0.15. The 3072x5504 portrait geometry additionally requires the maximum Green mid-band correlation to stay at or below 0.06. Its public score is the minimum normalized gate margin and crosses at 1.0.

The rule retained all 37 inferred candidates, including all seven official pixel positives. It accepted none of 17,417 exposed COCO, Open Images, and Spaces controls after the portrait alias gate. Because those controls informed the final gate, they do not certify the operating point. The constants were then challenged through the actual runtime on a frozen, decoded-pixel-deduped local holdout: 2,000 COCO val2017 images excluded from the earlier 3,000-image large color-phase challenge plus 637 unique Picsum controls deduplicated from 1,000 files. Four large geometries and four resampling kernels were balanced. The result was 0 of 2,637 detections, with maximum score 0.0592777965; the 95% zero-error upper bound is about 0.114%. These source collections existed before the freeze and supported other experiments, so the result is feature-unseen rather than a fresh-acquisition holdout.

The post-freeze Open Images acquisition was too slow to complete the planned 3,000-image challenge, but its completed subset provides an independent source audit. Incomplete .aria2 files, all prior Open Images decoded-pixel hashes, and duplicates were excluded before scoring, leaving 41 controls whose IDs and URLs had been frozen after the constants. The unchanged runtime accepted 0/41; the maximum score was 0.4083013324. This small audit supports the source boundary but does not replace the 2,637-control feature-unseen interval.

A prespecified post-freeze codec stress test re-encoded the seven official positives at unchanged dimensions. Native detection was 7/7, while JPEG-95 and JPEG-90 were each 0/7. Thresholds were not refitted. Large-v1 is therefore a native-pixel and lossless-copy detector, not a lossy-transcode detector.

The old fixed-only rule still crosses on four locally saved rows representing three unique pixel groups. One group is an Adobe file whose C2PA explicitly reports Google SynthID present. The other two remain queued for an official Gemini verdict and are not counted as clean negatives. The runtime branch is therefore positive-only and does not turn a miss into an absence claim.

Production routing now selects the fixed expert from 1 through 10 megapixels and the large expert above 10 through 18 megapixels when both sides are at least 2048 pixels. The score exporter emits fixed, registered, and large observations separately, preventing the large result from being mislabeled as a fixed score. The router accepts a registered or large threshold crossing only within that expert's measured scope and otherwise abstains.

2026-08-14: final OpenAI local rescue checks

An assertion-contrast probe at the dominant exact 1254-square geometry tested RGB, OKLab, and cube-root-LMS period-8 channel directions without treating missing assertion as a clean label. Validation selected srgb:p8:c2, accepting 45 of 259 asserted rows and zero of 13 same-provider plus three other-provider rows. The locked test accepted 39 of 233 asserted rows, zero of 15 same-provider rows, and one of three other-provider rows. The all-row audit accepted 185 of 992 asserted rows, one of 55 same-provider rows without an assertion, and one of 13 other-provider rows. A safe boundary above either wider control maximum retained only a small minority of the asserted cohort. Exact geometry therefore does not turn the weak OpenAI component into a production detector.

An independent published-decoder veto was then tested with the official InvisMark checkpoint. The downloaded checkpoint hash was fixed before use, and its bundled self-test reached 0.997 mean confidence with 0.97 recovered-bit accuracy. In contrast, OpenAI DTCWT hits, OpenAI misses, exact-1254 hits and misses, same-provider controls, Google, Microsoft, Canva, and unrelated controls all had median confidence near 0.18. The decoder is working, but its watermark family is absent from these cohorts; it cannot distinguish SynthID-bearing OpenAI pixels from renderer confounds.

Two representation-level missing-codeword hypotheses were also rejected. DTCWT modulus computed before spatial period folding selected off-diagonal phase coherence and retained 5 of 158 locked-test OpenAI rows versus zero of 135 controls, but zero of 40 fresh OpenAI rows. Whitening the period-8 feature space and clustering up to 16 codewords retained 5 of 158 and zero controls, then one of 40 fresh OpenAI rows and one of 50 fresh controls. Neither captured the known oracle-positive false negative. These were the remaining prespecified phase, wavelet, codeword, and external-decoder variants; none earns a local runtime route.

2026-08-15: OpenAI high-pass patch-bag challenge

One prepared but previously unexecuted content-dependent hypothesis aggregated the distribution of nine normalized high-pass patch embeddings per image. Its input contains luminance, Red-minus-Green, and Blue-minus-Yellow residuals but no filenames, metadata, dimensions, or RGB thumbnails. Mean, standard deviation, minimum, and maximum pooling allow spatially varying evidence without requiring one global phase codeword.

The validation-selected epoch accepted 285 of 1,280 OpenAI rows above every validation control. With weights and threshold frozen, the already exposed development test accepted 324 of 1,364 OpenAI rows, 3 of 1,373 difficult controls, and 1 of 2,000 COCO controls. The later August temporal cohort was not used by this model or threshold: native scoring accepted 120 of 486 OpenAI rows, 8 of 504 general controls, and 5 of 33 same-provider rows without a watermark assertion. Assertion enrichment over the same-provider stratum was not independently significant (p = 0.151, one-sided exact test). After decoded-pixel grouping the counts were 107/417, 4/453, and 4/30. Requiring the same unchanged score after JPEG-95 retained only 1/486 OpenAI rows and no controls.

None of the 120 native OpenAI hits overlapped the frozen signed period-8 hits; four overlapped the nonlinear period-4 candidate. This supplies new evidence for multiple complementary OpenAI raster components. It does not supply a production watermark decoder: same-provider non-assertion rows crossed at a high rate, general-control specificity failed, and ordinary JPEG destroyed all but one positive. A post-hoc threshold above the temporal control maximum would retain 19 unique OpenAI rows, but that threshold has observed the holdout and is not a new expert. The official OpenAI verifier remains the provider-wide pixel route.

2026-08-14: production OpenAI verifier boundary

OpenAI's official Content Provenance API now supplies the production-grade OpenAI pixel verdict that the local experiments could not justify. The runtime integration is deliberately a separate verify-openai-synthid command, never an implicit identify call. It requires the independent verify extra, OPENAI_API_KEY, endpoint access, and --acknowledge-upload.

The implementation establishes metadata independence before the request. It computes a decoded RGBA fingerprint, strips AI provenance metadata to a temporary copy, verifies that no AI markers survived and that the format and pixel fingerprint stayed identical, enforces the documented 50 MiB limit, and uploads only the temporary PNG, JPEG, or WebP. It parses exactly one synthid entry and deliberately ignores the independent C2PA outcome. C2PA-only positives, missing or duplicate SynthID fields, altered pixels, surviving metadata, unsupported formats, and 400/404/429 failures are covered by mocked tests. No credentialed API request was made during implementation because no API key was available.

A separate live smoke used OpenAI's public web verifier after running the exact production metadata-stripping and decoded-RGBA equality checks. Two OpenAI images at opposite 3:2 orientations returned SynthID detected and Content Credentials not detected. A Google SynthID oracle-positive control and a COCO natural control returned OpenAI SynthID not detected, again with Content Credentials absent. This confirms pixel-only, provider-specific semantics for the integration, but four fixed cases do not estimate statistical error rates and do not constitute a live SDK/API transport test.

On 2026-08-15, one separately preselected OpenAI-positive PNG completed the credentialed SDK path after the same metadata removal and decoded-pixel equality checks. The API returned an independent SynthID detected result with nullable model and generation-time fields, and the command completed without a retry. This closes the first positive PNG transport smoke only. It does not qualify JPEG, WebP, negative cases, operational rate limits, retention behavior, or local-detector accuracy.

This closes the production OpenAI detection surface through an official remote backend, not by relabeling the incomplete local period-8 expert. It does not expand the research oracle authority. The endpoint is not eligible for Zero Data Retention, and its documentation prohibits repeated reverse-engineering or evasion queries, so it cannot supply adaptive labels for detector or remover optimization without separate authorization.

2026-08-15: periodic codeword replacement and its measurement limit

Suppressing the estimated periodic residual with independent noise spends quality without reaching the decoder: at a strong budget the OKLab a/b branch moved the Gemini score by a median of -0.022 at 32.6 dB, because independent per-pixel noise averages out under the fold. Replacing the estimated period-16 residual with a magnitude-matched random-phase codeword instead moved all four Gemini positives below the frozen local threshold, with a median score of 0.050 against 0.533 at 53.5 dB and SSIM 0.9982. Adding non-periodic noise on top only cost 7 dB and changed nothing.

Both providers show a threshold, not a slope. Gemini keeps all four positives at replacement strength 0.75 and loses all four at 1.0; the OpenAI period-8 proxy keeps 9/16 at 0.25, 1/16 at 0.50, and none at 0.75. Random injection without the subtraction left 1/16, so the mechanism is subtract + randomize, not added energy.

Radial banding of the folded residual put the Gemini carrier in the middle band: 0.125-0.40 cycles/pixel gives a median -0.209 at 56.5 dB, against -0.009 above 0.40. The OpenAI proxy weights the upper band instead.

The follow-up question, whether the signal lives in the comb lines or in the broadband content between them, cannot be answered by the local score at all. fold_residual_template averages every tile, which is a comb filter: only frequencies that are integer multiples of 1/16 cycles per pixel survive, and everything between the lines is annihilated before the template is applied. probe_comb_vs_broadband.py measures the consequence rather than asserting it. Inside the same 0.125-0.40 annulus and at an identical residual RMS, randomizing only the comb bins moved the four Gemini scores by -0.177 to -0.519, while randomizing only the bins between them moved them by -0.0007 to -0.0012. A local null on the broadband branch is therefore guaranteed by the readout and is not evidence about the watermark.

That makes the two branches an oracle question. A frozen batch of four such triples is recorded in .local-eval/synthid/prc-oklab-attack-2026-08-15/comb-vs-broadband-oracle-manifest.json. Its first attempt established a control rule that the earlier runs had missed: the tracked originals carry a C2PA caBX chunk, and Google answered from it without ever reaching the pixel path, so an original is not a valid control for pixel-only candidates. Every arm is now re-encoded from decoded RGB with the chunk dropped and decoded pixels asserted identical. The re-encoded source then returned a pixel-path positive, which qualifies the control. The broadband candidate returned a temporary service limit with a two-hour retry, so no candidate verdict exists yet; an unavailable answer is not a negative.

2026-08-15: the local score is not the decoder margin

One group was then completed inside a single account, so that account tier and request routing stay fixed across its arms; spreading arms of one comparison across accounts would make a negative indistinguishable from a different backend. All four arms returned a Google positive:

Arm	Local score	PSNR	Google
source	0.416	-	detected
comb, equal RMS	0.239	53.97 dB	detected
broadband, equal RMS	0.415	54.27 dB	detected
full period-16 replacement at 0.95	0.087	52.56 dB	detected

The last row is the result that matters. That candidate sits well below the frozen local threshold of 0.174, which is the operating point the removal gate would have used, and Google still reports the watermark. A local score at or near zero therefore certifies nothing about the provider decoder, and no removal claim may rest on it. The equal-RMS pair adds that neither a comb-targeted nor a broadband-targeted perturbation at roughly 54 dB moves the verdict, so this run does not locate the signal either; it bounds what the local readout can say.

A replication attempt on a second image from a third account never produced a verdict: that account hit the service limit on its first request, while another had just served four. The limit therefore does not partition cleanly per account, and holding several accounts does not buy proportionally more oracle capacity. Plan oracle work around one window at a time and spend it on the arms that decide something.

Querying stopped here rather than escalating the perturbation budget until the verdict flipped. Searching for the breaking point by repeated queries is exactly the adaptive use of a provider oracle that this plan rules out, and the sham control was not spent because it only interprets a negative, and no arm returned one.

2026-08-15: the same answer from the OpenAI verifier

The equivalent frozen batch went through the official OpenAI endpoint on verify-openai-synthid, two OpenAI positives in a source and replace-chroma pair each, four calls, one pass. Every arm returned detected with metadata stripped, decoded pixels preserved, and the C2PA outcome ignored. The candidates sit at 54.13 dB and 53.97 dB with residual RMS near 0.50, the same perturbation class that empties the local proxy.

Two providers, two independent verifiers, one conclusion: a periodic-codeword replacement that drives the local score to zero does not remove what either decoder reads. The removal program has no evidence of pixel-only removal for either provider, and the local score cannot serve as its gate.

A third OpenAI positive was excluded before any call. Every periodic variant on it produced a residual RMS of exactly zero and a periodic energy ratio of exactly 1.0, so the attack changed nothing: it can only redistribute periodic energy it estimated, and it estimated none. That file is nevertheless recorded in the manifest as openai-verify: SynthID detected. An estimator that finds no periodic residual in an image the provider verifier flags is direct evidence that the period-16 and period-8 grids do not carry the whole signal, and a residual RMS of exactly zero is the signature to check for before reading any attack result as a null.

This batch was a single verification pass over frozen files. It is not an adaptive label source, and the endpoint's prohibition on repeated reverse-engineering or evasion queries stands: no candidate was adjusted and re-submitted, and no budget was escalated against the API.

A wider batch followed the same rule, sixteen pairs frozen in openai-oracle-batch-v2-manifest.json before the first call. Their perturbation budgets are not chosen, they fall out of each image's own estimated residual, and they span 37.9 to 57.5 dB. The API rate limit stopped the run after fifteen calls with a one-hour retry, leaving seven complete pairs plus one source. Every arm returned detected, including candidates as aggressive as 44.2 dB at residual RMS 1.58, roughly three times the perturbation that empties the local proxy.

The running total across both providers is nine source and candidate pairs and zero verdict changes. The strongest frozen candidates, near 38 dB, are still queued behind the rate limit; completing them finishes one pass rather than starting a new attempt.

2026-08-15: what the periodic features actually separate

The detector question needs no oracle at all. Both providers ship images from the same pipeline with and without a watermark assertion, identical claim generator and action list except for watermarked.unbound, so signed provenance supplies the control group directly. probe_render_fingerprint.py scores five groups with both shipped features.

Group	n	Google feature accepts	OpenAI feature accepts
google watermarked	300	83.3%	0.0%
google, same pipeline, no assertion	70	58.6%	0.0%
openai watermarked	300	2.3%	2.3%
openai, same pipeline, no assertion	287	0.7%	4.5%
unrelated control	299	7.4%	1.0%

Two results, and they land differently.

The OpenAI period-8 expert has no detection power on the general corpus. It accepts 2.3% of watermarked rows and 4.5% of same-pipeline rows without an assertion, so the difference is absent and its sign is inverted (p = 0.956 one-sided for the intended direction). The sixteen rows the frozen proxy "clears" are a selected subset, not evidence of a working expert, and this branch must not reach a shipped universal detector.

The Google period-16 feature does separate, but mostly by provider rather than by watermark. It accepts 83.3% of watermarked Google rows against 7.4% of unrelated controls, yet still accepts 58.6% of same-pipeline rows carrying no watermark assertion. The watermarked-versus-no-assertion gap is real (p = 1.6e-05, medians 0.513 against 0.206), so the feature is not purely a render fingerprint, but a majority of its accepts on that generator's own non-asserted output makes it a provider detector with a watermark-correlated component, not a watermark detector.

One caveat bounds both readings: a missing C2PA watermark assertion is not proof that no watermark was embedded. The no-assertion group is a lower bound on the false-positive rate, not a measured one.

Neither feature transfers across providers, 0.0% and 2.3% in the off-diagonal cells. That, and not the removal result, is the obstacle to a universal detector: there is currently no shipped feature that reads a watermark rather than a generator.

2026-08-15: a patch-local feature that beats the OpenAI branch

The evaluation set holds the confounds fixed rather than trusting them: one claim generator per provider, PNG only, and every control matched to positives of the exact same pixel dimensions, so size, codec and generator cannot carry the label. Folds are taken over matched strata, content hashes are deduplicated before assignment, and both the ridge strength and the feature family are chosen inside the training folds. That yields 64 Google groups over 301 images and 231 OpenAI groups over 1,081.

The features are patch-local and every one is a ratio: normalized radial band fractions, comb-to-background contrast at both periods, normalized periodic autocorrelation, and the residual's own small-lag autocovariance grid, in luminance and two opponent planes. Nothing is folded, so nothing is comb-blind.

Provider	Shipped AUC	Patch-local AUC	Delta, 95% CI over groups
openai	0.510	0.605	+0.095 [+0.045, +0.147]
google	0.665	0.629	-0.036 [-0.122, +0.050]

The OpenAI result is the first local expert on that provider with measurable detection power, against a shipped branch that sits exactly at chance. Twenty group-preserving label permutations put the null at 0.492 ± 0.020 with a maximum of 0.525, so the new score clears its own null by more than five standard deviations and by more than the largest permutation.

The Google branch is not improved and the shipped feature stands. A full-sample ablation showed the band subset alone reaching 0.676 there, above the shipped 0.665, but nested selection did not reproduce it, so that number was selection noise on the evaluation set and is recorded here only as the trap it is.

Two limits bound the claim. Sensitivity at ten percent control acceptance is 0.147 against the shipped 0.143, so the gain lives in the middle of the ROC and not at the precision-first operating point a shipped verdict needs. And a model fitted on one provider scores the other at 0.535 and 0.469, so this is a better provider-specific expert, not a universal detector.

An ablation locates the OpenAI signal in the normalized spectral bands, 0.608 alone against 0.571 for the autocovariance grid, 0.508 for comb contrast and 0.460 for the periodic lags. The two periodic families, which are what the shipped experts measure, contribute nothing on that provider. The across-patch spread block currently covers only the luminance channel, which is a defect in the feature vector rather than a finding; it scores at chance either way.

2026-08-15: the Google branch is control-limited, not feature-limited

A second attempt targeted three named causes rather than the score. The strata were deepened to twelve positives each, 533 against 63 controls. The fold assignment was repeated over seven seeds with the out-of-fold ranks averaged, because a single assignment over 64 groups made the inner selection noisier than the effect it selects on. And the shipped score was offered to that selection as one candidate among the feature families, so a result below the baseline means the selection failed rather than that the baseline is unbeatable.

It did not improve. Global AUC came to 0.678 against the shipped 0.689, a delta of -0.010 with a group-bootstrapped interval of [-0.071, +0.046].

Sensitivity at ten percent control acceptance did rise, 0.240 to 0.296, and that is the single number a precision-first reading would have quoted. It is an artifact of one threshold. Partial AUC over the whole region below ten percent control acceptance moves the other way, 0.164 shipped against 0.092, so the model is worse everywhere in the operating region except at the exact point the first metric sampled. One threshold is not an operating point.

The binding constraint is the control count. Sixty-three controls put the bootstrap interval on any delta near ±0.06, so an improvement of two or three points cannot be demonstrated here whatever the features do, and further model search on this set is selection on the evaluation data rather than progress. Unblocking the Google branch means more images from that generator known to carry no watermark, not more features.

That is what a provider verifier would supply, and Google has no open route to one. Three candidates were checked against their own pages on 2026-08-15; the fetched evidence is archived under .local-eval/synthid/provider-access-2026-08-15/.

Vertex WatermarkVerificationModel is not a route. The class survives in vertexai.preview.vision_models with verify_image returning a decision field, but the whole generative-AI preview module is marked deprecated as of 2025-06-24 with removal scheduled 2026-06-24, a date already past, superseded by the Google Gen AI SDK. The documentation page vertex-ai/generative-ai/docs/image/verify-watermark returns HTTP 404, including its locale variants and the legacy cloud.google.com redirect. Whether the backing endpoint still serves is unverified, and no successor page was found. An earlier revision of this document named this the programmatic route; that was wrong.

Google Cloud's AI Content Detection API is real but answers a different question. It is documented at gemini-enterprise-agent-platform/models/ai-content-detection (last updated 2026-08-13) as Private Preview behind an allowlist intake form whose own title is go/detect-ai-content-allowlist. The page never uses the words SynthID or watermark. It describes a classifier that analyzes "pixel-level artifacts, noise patterns, and spectral anomalies" and states that "Support for C2PA metadata detection is not included". So it decodes no watermark and no manifest: it is a probabilistic AI-generated-image classifier. It cannot supply the ground-truth negatives the Google branch is short of, though it would serve as an independent comparison baseline for our own detector.

The DeepMind SynthID Detector reads the watermark but is closed to us. It is a portal, not an API, and its early-tester form ANDs three requirements, the first being "You are an active journalist or verification professional", with a news-publisher corporate email and required News Publisher and Publisher URL fields. The words research, academic and developer do not appear on the form at all. Research access is excluded by construction rather than by refusal.

The Gemini app therefore stays the only working route, at roughly two checks per account per two hours. The Google branch's control shortage has no near-term provider-labelled fix.

2026-08-15: the two providers are architecturally different (S4)

Cross-image correlation of the folded residual, unit-normalized, over the matched sets plus 200 unrelated controls. Three pairings are needed, not one: a high positive-to-positive correlation alone cannot tell a watermark carrier from a generator fingerprint, because the controls come off the same generator.

Provider	Representation	pos-pos	ctl-ctl	pos-ctl	pos-unrelated	chance
google	tile16	+0.326	+0.158	+0.224	+0.023	0.036
google	tile8	+0.420	+0.223	+0.304	+0.035	0.072
openai	tile16	+0.032	+0.024	+0.027	+0.008	0.036
openai	tile8	+0.035	+0.031	+0.033	+0.009	0.072

Google images share a fixed, phase-coherent pattern at both geometries, an order of magnitude above chance, and unrelated images do not carry it at all. OpenAI images share nothing: every cell sits at or below the chance floor, which is what a content-dependent post-hoc neural encoder produces and what the SynthID-Image paper describes. The comb experts work on one provider and not the other because the providers are not doing the same thing.

The Google numbers say something sharper than "positives correlate more". A single shared pattern present in two classes at different amplitudes gives a cross-correlation at the geometric mean of the two within-class values, whereas two distinct components would fall below it. Observed against predicted: 0.224 against 0.227 for tile16, 0.304 against 0.306 for tile8, ratios of 0.985 and 0.993. So it is one pattern, not a watermark carrier layered on a separate generator fingerprint, and the controls carry that same pattern at roughly half the amplitude of the positives.

Three readings survive that, and L1 decides between them: the controls are watermarked after all and the label is wrong; the mark modulates the amplitude of a fixed carrier the generator always emits; or the carrier is a generator artifact whose strength happens to track the assertion. The first would also explain the Google feature's 58.6% acceptance on controls without any of it being a false positive.

The magnitude-only rows are reported in shared-carrier-report.json but carry little: discarding phase leaves the generic image spectrum, and unrelated photographs already correlate at +0.348 there, so the floor is too high for the measure to separate anything.

2026-08-15: the Google tile space is exhausted

S4's geometry suggests an estimator. If the classes differ by the amplitude of one shared pattern, a discriminant whitened by the within-class covariance should beat a plain cosine against a template, which weights all 768 tile dimensions alike. detector_tile_lda.py fits exactly that, in the same space, with the family and penalty chosen inside training folds and seven repeated fold assignments.

It does not. Google scores 0.644 against the shipped 0.689, a delta of -0.044 with interval [-0.090, +0.004].

The failure follows from the finding rather than contradicting it. When two classes carry ONE pattern at different amplitudes, the optimal statistic IS the projection onto that pattern, which the cosine already computes; whitening adds estimation variance over 64 controls and buys no direction that was not there. Read together with the earlier 124-dimensional attempt, this closes the question the two of them were confounding: the Google shortfall is not the metric and not the feature bag. The folded tile simply contains one usable degree of freedom, and the shipped score already spends it.

Any Google improvement must therefore come from outside the fold, from the content the comb filter destroys before the template is applied. That is S2, and S2 needs the provider oracle.

One reporting caveat: this run scored both providers against folded_template_score, which for OpenAI is the Google template applied to foreign images and not that provider's shipped expert. The OpenAI delta printed in detector-tile-lda.json is therefore against the wrong baseline. Against the right one the tile discriminant reaches 0.565 versus the period-8 expert's 0.510, which is still below the 0.605 the patch-local features already reach, so it is not an improvement on either provider.

2026-08-15: the implementation is sound, the date is innocent, the product is not

Four register entries closed offline in one pass.

M3, the shared implementation, holds. Every removal and detection result runs through folded_template_score, so a defect there would produce both at once and look like two findings agreeing. probe_m3_independent_fold.py re-derives the statistic from its definition over 40 real images, folding by explicit modulo accumulation and blurring with a hand-rolled separable Gaussian rather than the OpenCV call, so a shared bug has nowhere to hide. Maximum score gap 3.7e-4, median 1.3e-4, against scores spanning 0.01 to 0.85; the residue is the shipped path's float32 blur. The implementation is not the explanation.

L3, the date confound, is refuted. Month alone predicts the label at 0.524 on Google and 0.491 on OpenAI, both chance. Inside a single month the Google signal is not merely intact but stronger than pooled: 0.723 in 2026-06 over 44 controls and 336 positives, against 0.689 pooled. The OpenAI shipped expert stays at chance within every month, as it is everywhere.

R2 says the OpenAI result is not a geometry artifact, and adds a caution about Google. Restricted to one exact pixel size and pooled over 13 strata weighted by pairs, the patch-local feature scores 0.577 against the shipped 0.478, an edge of +0.099 that matches the +0.095 measured pooled. So the improvement survives the strongest available geometry control. Individual strata swing from 0.381 to 0.799, which is what 5 to 55 controls per stratum buys and is why the pooled figure is the one to quote. Google has only three usable strata and one dominates them, so its within-size figures, 0.767 patch and 0.763 shipped, are recorded as underpowered rather than as an improvement.

D1 is the finding that matters for the product, and it is bad. Held at one percent control acceptance, which with 63 and 229 controls means clearing every control in the set, the fraction of positives that can be called at all is: Google 0.045 shipped and 0.000 patch-local; OpenAI 0.025 shipped and 0.018 patch-local. Every headline AUC in this document lives in the middle of the ROC. At a precision-first threshold neither provider's detector can call more than a twentieth of true positives, and on Google the newer feature calls none.

That reframes the program. The gap between 0.605 AUC and a shippable verdict is not a matter of a few more points; nothing currently supports a positive claim at a defensible false-positive rate. Either the operating point moves, which means accepting a control acceptance far above one percent and saying so in the product surface, or the representation has to change enough to move the tail rather than the middle.

2026-08-15: U2, the detectors say "synthetic", not "watermarked"

Until now every control was a photograph or same-provider output. Neither can catch the failure that decides whether a positive verdict may be reported as a provenance claim: firing on an AI-generated image from a generator that embeds no SynthID. The corpus carries 223 such images with their own signed provenance, so the label is read from a manifest rather than inferred from a path.

Acceptance rates, each feature at its own threshold. The patch-local models have no published threshold, so each is cut where one in ten of its own provider's legitimate controls is already accepted.

Generator	n	Google shipped	OpenAI shipped	patch Google	patch OpenAI
Adobe Firefly	86	0.34	0.06	0.51	0.24
Microsoft Designer	20	0.20	0.00	0.40	0.25
Microsoft Responsible AI 1.0	55	0.00	0.24	0.31	0.20
PixelBin.io	35	0.31	0.00	0.37	0.09
fal-ai/seedvr	10	0.00	0.00	0.80	0.00

The Google shipped feature accepts 7.4% of unrelated photographs and 83.3% of watermarked Google images. It accepts 34% of Adobe Firefly output, which carries no SynthID. The patch-local models are worse: cut at a level that admits one in ten legitimate same-provider controls, they admit half of Firefly and four fifths of fal-ai.

No threshold repairs this. A detector that fires this often on foreign synthetic images is measuring generation, not the watermark, and a positive verdict from it cannot be reported as "SynthID detected" without asserting a provenance the evidence does not support. That is a correctness bound on the product surface, not a performance target: it constrains what a positive may be called, whatever the AUC.

It also bounds the reading of every earlier number here. The Google feature's 0.665 to 0.689 AUC against same-provider controls was already suspected of tracking the generator; these rows show it tracks synthesis more broadly than that, across vendors with unrelated pipelines.

2026-08-15: a high number that must not be quoted

Following U2, the features were asked to separate watermarked images from AI-generated images of other vendors, whose own provenance says they carry no SynthID. Folds keep each vendor whole, so no fold trains and tests on the same vendor.

Provider	vs foreign generators	vs same-provider controls
google	0.917	0.636
openai	0.806	0.610

The high column is confounded and is recorded here so that it is not later mistaken for a result. Separating a Google image from an Adobe Firefly image is easy because the generators differ, and the same features already score 0.917 and 0.806 at telling providers apart in every other experiment in this document. The comparison cannot isolate the watermark, because changing the control set also changed the generator.

The design flaw is not fixable by choosing better foreign vendors: any control drawn from a different generator carries generator identity along with the absence of the mark. The only control that varies the watermark while holding the generator fixed is same-generator output without the mark, which is the scarce set L1 is currently validating and the set M2 proposes to mint. Every unconfounded discrimination this program has measured sits at 0.61 to 0.69, and that band is the honest state of the detector.

2026-08-15: M1, the pipeline is not the problem

Every negative result here is ambiguous between two causes: the features cannot see watermarks, or SynthID is harder than the data allows. Embedding a watermark ourselves separates them. The ground truth is not a provenance assertion that may be wrong, and each carrier is its own control, so image and mark never straddle a fold.

invisible-watermark's dwtDct mark was embedded into 220 unrelated carriers at three amplitudes, the library's own and two attenuations, because a mark loud enough for a naive detector proves nothing about a signal at SynthID's level.

Strength	AUC	TPR at 10% FPR	median residual RMS
full	0.990	0.995	2.28
half	0.974	0.977	1.17
quarter	0.936	0.895	0.563

The last row is the one that matters. A residual RMS of 0.563 is the amplitude band this program has been measuring SynthID in all along: the estimated periodic residuals and every attack candidate sit between 0.45 and 0.6. At that amplitude the same feature extractor, the same folds and the same estimator reach 0.936, against 0.53 to 0.69 on real provider data.

So the machinery is sound at the relevant signal level, and the shortfall on real data is a property of the task rather than of the code. Two candidate causes survive, and they are not exclusive: SynthID is structured unlike a fixed-carrier mark, which S4 already showed directly for OpenAI, and the labels are heavily contaminated, which L1 is showing now.

The bound on this result: dwtDct applies one payload by one algorithm to every carrier, so all marked images share a systematic perturbation. That is an easier target than a content-adaptive neural encoder. M1 establishes that the pipeline finds a fixed-structure watermark at SynthID's amplitude, not that it would find any watermark. Repeating it with a neural watermarker is the next step, and it would close the remaining ambiguity.

2026-08-15: what property of a mark actually defeats these features

M1 left one ambiguity: it found a fixed-structure mark and missed SynthID, and content-adaptivity was the obvious suspect. Synthetic marks separate the properties, all scaled to the same residual RMS so amplitude explains nothing.

Mark	AUC	median residual RMS
fixed, one pattern for every carrier	0.670	0.626
keyed, a different pattern per carrier, same band	0.669	0.626
keyed and perceptually masked by local texture	0.546	0.577

Content-adaptivity is not the obstacle. Fixed and keyed are indistinguishable, 0.670 against 0.669, because the band features read the shared spectral signature rather than the pattern. Perceptual masking is the obstacle, and it costs three quarters of the separation.

That places every number in this document on one scale. At a residual RMS near 0.6: dwtDct scores 0.936, being structured and unmasked; diffuse band noise scores 0.670; masked diffuse noise scores 0.546. Real SynthID scores 0.53 to 0.69. The provider mark behaves like a masked, diffuse, mid-band perturbation, which is what a competent perceptual encoder is built to be, and our 0.605 is at the ceiling of what global hand-crafted statistics reach against one, not short of it.

Undoing the mask does not help. Dividing the residual by the same local-activity estimate before extraction scores 0.538 against 0.546 ordinary. The reason is structural rather than a tuning failure: the band features are energy fractions and therefore scale-invariant, so a mask that is nearly constant across a patch already cancels. The loss comes from the correlation masking creates, not the scaling. A mark whose amplitude tracks local texture is strongest exactly where the image's own residual is strongest, so its local signal-to-noise is flat and equal to its global value, and no region remains where it stands out. Statistics of this family live on such regions.

The consequence for the program is concrete. Beating a perceptually masked mark requires knowing its pattern, which is a decoder with the key, not a better description of the residual. Absent the key, a local detector is bounded near where ours already sits, and the remaining levers are the label quality that L1 is measuring and a learned decoder trained on genuinely clean pairs.

A defect found and fixed in passing, recorded because the first version of this test returned a plausible number: the demasking probe originally rebuilt an RGB image from the corrected residual with a hand-written inverse of the opponent transform whose round-trip error was 15 grey levels, against marks of half a level. It scored 0.529 and looked like a clean negative. The residual is now handed to the extractor directly, and the probe asserts that a supplied residual reproduces the extractor's own features exactly before it measures anything.

2026-08-15: the Google label rule is refuted, and the Google feature is blind

The M2 pilot was meant to mint unwatermarked pairs with Imagen's addWatermark. It failed at the first step and produced something more decisive than it was designed to.

Imagen is no longer in the model garden. All of imagen-4.0-generate-001, imagen-3.0-generate-002 and imagegeneration@006 return 404, and a listing of the 127 available publisher models contains no imagen-* entry. Image generation now runs on gemini-2.5-flash-image and the gemini-3.x-*-image family, and those reject the parameter outright: Unknown name "addWatermark" at 'generation_config': Cannot find field. There is no watermark toggle on any current Google generation path, so M2 as designed cannot be built.

The Gemini API documentation states the policy directly, twice, rather than leaving it to inference: "All generated images include a SynthID watermark." Its optional-configuration section lists output format, aspect ratio and image size, and nothing else. Generated images are therefore watermarked by construction.

Two results follow, and both are read off documentation plus four generated images rather than off an oracle.

The Google feature does not see SynthID on current output. Four freshly generated images scored 0.018, 0.103, 0.054 and 0.071 on folded_template_score, a median of 0.063, against a corpus median of 0.51 for watermarked rows and 0.01 for unrelated photographs. The feature is not over-firing on unwatermarked images; it fails to fire on watermarked ones. Taken with the U2 result, it detects a generator era rather than a watermark, and that is now demonstrated rather than suspected.

L1's refutation extends to Google. Every one of those four images carries a C2PA manifest from the same claim generator as the corpus positives, "Google C2PA Core Generator Library", and none of them asserts watermarked.unbound; their action lists are c2pa.created and c2pa.edited only. Google's own API emits a watermarked image with no watermark assertion. The rule that a missing assertion means a missing watermark is therefore refuted for both providers, and the 70 Google controls are unreliable in exactly the way the 287 OpenAI controls were measured to be. Every Google AUC in this document rests on that label.

A method note. The first attempt to settle this empirically was going to upload a generated image to the Gemini app verifier. The composer renamed the attachment from .png to .jpg: the interface re-encodes uploads. A not_detected verdict after re-encoding would have been uninterpretable, and the query was abandoned rather than spent. Today's earlier oracle runs went through the same interface and returned detected, so the mark survived re-encoding there, but no negative verdict obtained through that surface should be trusted without controlling for it.

2026-08-15: what the corpus labels actually mean, and what that invalidates

Earlier entries in this log, written the same day, describe both providers' label as the watermarked.unbound action. That is wrong for Google and the error propagated into every Google experiment here. The corpus was counted directly.

watermarked.unbound appears in 0 of 4,780 Google rows and 7,427 of 7,731 OpenAI rows. It is an OpenAI field. Google's synthid_from_provenance flag is driven instead by synthid_watermark, "present according to Google LLC provenance", which is set from the presence of a Google generation claim.

The two classes therefore differ by something other than a watermark:

	Google positives, 4,710	Google "controls", 70
claim generator	Google C2PA Core Generator Library	Google C2PA Core Generator Library
issuer	Google LLC	Google LLC
actions	created, converted, edited, opened and similar, always with created	converted, edited, opened, never created
platform	Google (Gemini / Imagen)	C2PA signer: Google LLC (no known AI generator named)

The controls are images Google edited or converted but did not create. So every Google number in this document, the 0.665 to 0.689 shipped AUC, the patch-local comparisons, the within-month and within-size splits, measured Google-generated against Google-edited. That is a synthetic-versus-photographic task, not a watermark task, and it explains U2 without any further hypothesis: a classifier trained that way fires on Adobe Firefly because Firefly output is synthetic.

There is no Google watermark control set in this corpus and there never was.

Worse for the program, one cannot be collected. Google's documentation states that all images generated by its image models carry a SynthID watermark, and the addWatermark toggle that could produce an exception exists only on Imagen, which returns 404 for every documented model id on an accessible project. The same documentation adds a restriction that rules out the cleanest experiment even with access: "You can use the seed field to get deterministic output only when this field is set to false", so an identical image with and without the mark cannot be generated at all.

For OpenAI the label is the right field but the negative class is contaminated, measured at 20 of the first 28 non-asserted rows carrying a watermark.

What survives from the day's work is what did not depend on these labels: M1 and M1b, which used marks embedded here with known ground truth, and M3. The rest needs redoing on a task that is stated correctly first.

2026-08-15: the periodic lattice is JPEG, not SynthID

Classifier work was abandoned here for signal characterization. A discriminative model finds whatever separates its classes, and this program has now measured that what separates them is the generator and its era, so every trained number carries that confound inside it. The question was changed to what the watermark does to pixels.

Without a clean copy of the same image the perturbation is isolated by averaging: a content-independent component accumulates over hundreds of images while content cancels. Three groups make the comparison specific. Photographs carry neither mark nor synthesis. Foreign-generator images are synthetic and carry no SynthID, so subtracting them removes what is generic to generation and leaves what is particular to Google.

The difference looked, at first, like a discovery. Google minus foreign concentrates on the period-8 lattice far beyond chance: the lattice holds 0.024% of bins and takes 2.7% of the difference in luminance, 10.8% in red-green and 7.8% in blue-yellow, enrichments of 112x, 441x and 318x. The strongest individual bins sit at one eighth and three eighths of a cycle per pixel, on the same lattice.

Period 8 is also the JPEG block size, and the control settles it. Measuring each group's own lattice concentration rather than a difference:

Group	L	R-G	B-Y
corpus Google	17.8	117.4	70.1
foreign generators	18.9	73.1	68.9
photographs	9.1	30.4	25.5
fresh Google from the API	3.4	9.5	6.2

The fresh images are watermarked by documented policy and have never been through a lossy codec, and they carry the least lattice structure of any group. If the lattice were the watermark they would carry the most. Foreign generators carry nearly as much as Google. The ordering tracks compression history.

So the periodic lattice this program has been measuring, attacking and building experts on since the beginning is block-transform structure, not SynthID. That one fact retro-explains the rest without further hypotheses: why the shipped detector scores 0.51 on corpus positives and 0.063 on fresh output from the same provider, why it fires on Adobe Firefly, and why S4 found a shared fixed pattern among Google images that the controls also carried.

It also reinterprets the removal programme. Nine oracle pairs changed no verdict at up to three times the perturbation that empties the local score, and the reason is now plain: the attacks removed and replaced compression structure. The watermark was never touched, so there was nothing for a provider decoder to lose.

2026-08-16: the large expert detects fresh out-of-distribution output

The previous entry concluded that the periodic structure this programme has been measuring is block-transform artifact rather than SynthID. That conclusion was drawn from 1024-square API images and generalized too far. It is corrected here.

Reading the earlier log supplied the constraint that prompted the check. The 2026-08-14 entry records large-v1 validated against seven official Google pixel verdicts and accepting none of 2,637 feature-unseen controls, and it records a codec stress test in which JPEG-95 and JPEG-90 each took detection from seven of seven to zero of seven. A compression artifact would be reinforced by re-encoding, not destroyed by it.

gemini-3.1-flash-image on the global endpoint accepts imageConfig.imageSize and returns 5632x3072, 17.3 megapixels, inside the large expert's routed scope and one of the geometries it was calibrated against. Eleven such images were generated, never written through a lossy codec, and watermarked by the documented policy that all generated images carry SynthID.

The shipped runtime detected eight of eleven. The crossing scores cluster just above the gate at 1.036 to 1.061; the three misses returned 0.637, 0.568 and 0.184 and were reported indeterminate rather than negative.

That model did not exist when the expert was calibrated, and these files have no codec history, so this is generalization across both model generation and pipeline. large-v1 is a working local SynthID detector for large images at roughly 73% sensitivity on out-of-distribution positives.

The corrected reading of the two experts:

• the large expert measures something real that survives a new model generation and the absence of compression;
• the fixed period-16 expert scores a median of 0.063 on fresh 1024-square output against 0.51 on the corpus, and the corpus small-image chroma lattice is ten times the fresh one at matched size with luminance nearly equal, which is the signature of 4:2:0 chroma subsampling. Its corpus performance is therefore still best explained by compression history, but that explanation is now specific to the small-image expert rather than to the periodic family.

What remains genuinely open is whether the fixed expert is confounded or merely out of regime: 1024 squares may be below the size at which this signal is recoverable at all, which the large expert's scope already hints at.

2026-08-16: a pre-registered test that failed, and why it could not have worked

The mid-band chroma lead from nine images was carried into a confirmation run on fresh data the hypothesis had not seen. It is recorded here because a pre-registered test is worth nothing if only its successes are written down.

Against 223 foreign-generated controls, band 4 in the opponent channels came back at z = -8.33 and z = -5.69 on 150 fresh OpenAI images: enormous, and pointing the opposite way to the lead. The exploratory scan behaves the same, with bands 0, 1 and 6 and the lag-one autocorrelation all separating at |z| > 6.

Two reasons this is not evidence about a watermark. The lead came from Google images and only the OpenAI arm reached usable size, so the populations differ. More fundamentally, the control differs in generator, so a large z here says gpt-image at low quality has different spectral statistics from Firefly and Microsoft output, which was never in doubt. This is the same confound identified in the foreign-generator classification test, and averaging does not escape it: without an unwatermarked control from the same generator, no amount of n converts a generator difference into a watermark difference.

The Google arm stands at ten 1024-square images, too few to test, and the same objection would apply to it.

The practical consequence points at the day's other result. The signal that is demonstrably present is the one large-v1 reads at 17 megapixels, and every OpenAI image collected here is 1024 square, which on the Google side is the regime where nothing is recoverable. The next OpenAI attempt should be made at large geometry rather than with more small images.

2026-08-16: a seven-pixel shift, and what large-v1 actually reads

The crop ladder was run to find the size at which the signal stops being recoverable. It found no such size, and answered a different question instead.

Cropping preserves pixels and lattice phase where resampling would not, so crops of the eleven fresh 17-megapixel positives isolate size from every other variable. Two ladders were taken, one from the origin and one from an offset of seven pixels, which is deliberately not a multiple of the tile period. Components are medians over the eleven images, folded whole rather than per window.

Crop	Aligned fixed	Aligned R-G	Shifted fixed	Shifted R-G
5632x3072, 17.3 MP	0.338	0.962	-	-
4096x2560, 10.5 MP	0.342	0.962	0.044	-0.111
2048x2048, 4.2 MP	0.335	0.963	0.044	-0.104
1024x1024, 1.0 MP	0.366	0.959	0.039	-0.081
512x512, 0.3 MP	0.350	0.937	0.046	-0.068

Two facts, and the second settles the programme's central question.

The structure does not decay with size at all. It is as strong in a 512-square crop as in the full 17 megapixels, so the fixed expert's blindness on small images was never a size limit.

And it is phase-locked to the image origin. A seven-pixel shift takes the red-green agreement from 0.962 to -0.111 at every size.

That is disqualifying. The SynthID-Image paper reports 99.97% TPR in its "Spatial worst" category, aggressive crop combined with resize, evaluated at 512 squares, and Google's own page describes the mark as "designed to stand up to modifications like cropping". A structure destroyed by a seven-pixel crop cannot be a watermark whose published evaluation survives far worse.

So large-v1 reads a generation-pipeline artifact: origin-anchored, size-invariant, chroma-dominant, at the tile period. A tiled upsampling or decoding grid fixed to the image corner has exactly those properties. Every observation about it now follows without further hypotheses: seven official positives that all came from one pipeline family, none of 2,637 photographic controls which have no origin-locked lattice, eight of eleven fresh gemini-3.1-flash-image outputs from the same family, 0.063 on gemini-2.5-flash-image output from a different one, and destruction by JPEG, which imposes its own grid and breaks the phase agreement.

A method note that should outlive this entry: the shift test costs one line and would have settled this in a day at any point in the last three months. Any candidate local expert should be required to survive a non-aligned crop before it is calibrated, let alone shipped, because the mark it claims to detect provably does.

Also recorded, since it produced a contradiction worth explaining: the table above folds each crop whole, while large_image_components minimises its components over phase-aligned 2,048 windows. The two are not the same statistic, which is why the shipped runtime detected eight of eleven full images that this table's gate arithmetic rejects. The aligned-versus-shifted contrast does not depend on that difference.

2026-08-16: specificity measured against what the route can be confused by

Every control rate this project publishes was measured on photographs: 0/2,637 for the large branch, 0/5,993 and 0/2,366 for registered-v3, 0/41 on a post-freeze acquisition. Photographs are the easy negative, because they carry no origin-anchored lattice at all. Once the route is named a generation-pipeline detector, the negative that decides its worth is other generators' output.

The production entry point was run over the 223 corpus images whose C2PA claim generator is a non-Google vendor, so the cohort's composition is signed rather than assumed.

Claim generator	Detected	n	Rate
Adobe Firefly	21	87	0.241
Microsoft Designer	3	20	0.150
PixelBin.io	5	36	0.139
Microsoft Responsible AI Provenance	0	14	0.000
fal-ai/seedvr	0	10	0.000
Microsoft Responsible AI 1.0	0	56	0.000
total	29	223	0.130

The shipped runtime accepts about one in eight images from other generators, with a one-sided 95% lower bound of 0.095, and reaches 0.241 on Adobe Firefly. The highest foreign score was 3.01 against a threshold of 1.0, so these are not marginal crossings. Every acceptance came from registered-v3; the large branch was either unsupported or indeterminate on this cohort.

That is the number a user meets. A tool that answers "is this AI-generated content watermarked" will be pointed at AI-generated content, not at landscapes, and against that population the published photographic control rates do not describe its behaviour.

One measurement in the same run is reported but must not be read as a false-positive rate. A cohort of 300 corpus rows with neither provider nor claim generator was accepted at 0.090. Those rows are unlabelled rather than verified photographs and may contain AI images stripped of metadata, so the number is uninterpretable in either direction. The COCO and Open Images collections behind the published figures are not on disk, so the 0/2,637 result could not be reproduced here and is neither confirmed nor contradicted.

2026-08-16: registered-v3 is phase-locked too, so the whole runtime is

Every acceptance in the entry above came from registered-v3, the default ordinary-size route, and the seven-pixel shift test had only ever been run on large-v1 and the large branch's components. The same test was therefore run through the production entry point on the ordinary-size expert: the 29 foreign-generator detections, plus 12 Google provenance-signed positives chosen inside the registered geometry, each scored at diagonal crop offsets 0, 2, 4, and 8 with a 32-pixel margin on every side.

Cohort	n	Detected at 0	Detected at 2	Detected at 4	Detected at 8
Foreign-generator detections	28 in cropped geometry	28	0	27	25
Google provenance positives	12 (8 detected at 0)	8	0	8	8

One foreign image left the registered geometry under the crop and could not be shifted. Every one of the 36 detections that remained in geometry died at a two-pixel diagonal offset: maximum foreign score 0.779 and maximum Google score 0.311 against the 1.0 threshold, with medians of 0.0. Four of the twelve Google positives were indeterminate at every offset, a sensitivity miss rather than a phase effect.

The recovery at offsets four and eight is itself the signature. Forcing the period-16 fold on one Adobe Firefly detection gives 0.808 correlation at offset zero, -0.008 at offset one, about 0.14 at offset two, and back above 0.77 at offset four: the matched structure repeats every four diagonal pixels and is destroyed by anything in between. Period selection breaks as well; the canonical winner jumps from 16.000 at offset zero to 9.590 at offset two and the base statistic falls from 2.47 to -0.58.

A mark designed to survive cropping would score the same under a two-pixel crop; the published SynthID evaluation retains 99.97% TPR under aggressive crop and resize. Registered-v3 therefore reads the same origin-anchored, crop-destroyed lattice established for large-v1 in the seven-pixel entry, and reads it in both directions: on the foreign generators it fires because those pipelines also leave a sixteen-pixel origin-anchored structure that partially matches the frozen Google template, and on the Google positives it reads the same class of structure rather than a robust mark.

The S1 conclusion now covers every shipped periodic expert, not only the large branch: the entire local pixel route detects generation-pipeline lattice, and identify's experimental pipeline_lattice signal is the honest name for all of it. Any future expert must pass the two-pixel shift test before calibration, per the R0 precondition.

2026-08-16: hypothesis register audit

A full pass over the register against the recorded evidence, closing what the log already settles:

• M2 re-verified independently and settled. imagen-4.0-generate-001 and imagen-3.0-generate-002 return 404 in us-central1 and us-east1 on both GCP projects (gen-lang-client-0926942364 and raiw-cws-publish); gemini-2.5-flash-image rejects addWatermark at request parsing (Unknown name "addWatermark" at 'generation_config': Cannot find field), and the current Gemini API documentation states for both the Gemini-image ("Nano Banana") and Imagen paths that all generated images include a SynthID watermark, with no disable parameter documented. No current Google path can emit an unwatermarked image.
• S2 moves from open to partial: the equal-RMS broadband arm of the account-isolated Google oracle group remained detected, first direct evidence against the between-lines scope at the tested perturbation class.
• S3 gains the Google nested test, which did not improve on the shipped score (0.678 vs 0.689), so the Google branch stays control-limited.
• L3 is settled by a direct date-only classifier (AUC 0.524/0.491).
• M3 is settled by the independent fold reimplementation (3.7e-4).
• D1 was asked and measured: 0-4.5% of positives decidable at one percent control acceptance.
• The ranked backlog gains an execution-state table: H1 executed and closed as a watermark route (it locks onto the S1-refuted pipeline lattice), H2 rejected, H3 partially executed, H4 rejected as a standalone detector, H5 advanced as research confirmation only, H6/H7 blocked on the oracle-positive corpus, H8 not started.

Nothing in this pass reopens a closed hypothesis, and no status moved on an argument; every change cites a recorded measurement.

2026-08-17: era, vendor and codec audit of the patch-local bank

The power-spectrum bank cannot be blinded by translation, so the plain crop test that killed the phase-locked experts certifies nothing for it. The axes that can still separate a mark candidate from a pipeline signature are era (the statistic must fire on corpus-era June-July positives, fresh gemini-2.5-flash-image positives, and fresh gemini-3.1-flash-image 4K positives alike, since the documented policy watermarks all three), vendor (silence on C2PA-signed Firefly, Microsoft, PixelBin and fal-ai output), and codec (survival of JPEG-95 applied to both classes, which the published mark survives). probe_era_vendor_codec.py scored all 124 features over that grid: positives 533 corpus rows plus 10 fresh-2.5 plus 11 fresh-4K, negatives 218 foreign-generator images, 300 seeded COCO photographs and 100 exact-1024 Open Images photographs, three conditions each.

Result: zero features clear a minimum AUC of 0.55 over the 27 cells, zero clear 0.6, 0.65 or 0.7. The best worst-cell is 0.481. The strongest cells are the familiar ones, for example feature 47 at 0.9997 on corpus-native against photographs, and every such feature collapses below chance in some other era or codec cell. A ridge combination fitted on corpus-era positives transfers to the fresh eras at AUC 0.388 native and 0.178 after JPEG-95; fitted on fresh 2.5-flash it transfers back to the corpus era at 0.243. The shipped folded baseline repeats the phase-lock story per era: corpus 0.8845 native against foreign generators falls to 0.4479 under a random 1-15 px crop, fresh 4K falls from 0.859 to 0.4508. The one weak exception is the ten fresh 2.5-flash images at 0.6114 under crop, which is also the set whose negatives are JPEG-stored while the positives are PNG, so the codec asymmetry is untested there and the sample is ten images.

What this settles: every separating statistic in the hand-crafted bank is era-, vendor- or codec-bound, for the Google side measured directly and for the OpenAI side already implied by the reversed pre-registered band test of 2026-08-16. Combined with the M1b masked-mark ceiling this closes the hand-crafted branch as a SynthID route for both providers. The one untested local route left is S5, a learned nonlinear model on residual patches, which now carries the additional requirement demonstrated here: leave-one-era-out transfer is the acceptance gate, because within-era separation is proven to be pipeline separation.

2026-08-17: S5 first slice, learned patch detector under the era gate

probe_s5_patch_conv.py trains a 4-layer convolutional model on opponent-plane Gaussian residual patches, eight unaligned random 128-px patches per image, so no phase anchor to the image origin can exist by construction. Positives span the eras available on disk: 535 corpus-era Google rows, 10 fresh gemini-2.5-flash-image, 11 fresh gemini-3.1-flash-image 4K, 853 corpus-era OpenAI rows and 151 fresh gpt-image-1. Negatives are 223 C2PA-signed foreign generator images and 400 photographs (300 seeded COCO, 100 exact-1024 Open Images). Each provider's model is trained on all eras but one and evaluated on the held-out era, native, after JPEG-95 on both classes, and after a two-pixel shift; the shift column equals the native one as designed, confirming the construction. Three image-level label permutations put the null at 0.552, 0.553, 0.593 for Google and 0.540, 0.566, 0.554 for OpenAI.

The verdict repeats the U2 result for a learned model. Against photographs the held-out eras score 0.711, 0.599 and 0.843 for Google and 0.791 and 0.767 for OpenAI. Against foreign generators the same models score 0.442, 0.153 and 0.486 for Google and 0.455 and 0.153 for OpenAI: at or below the permutation null, with the fresh eras driven below chance, meaning the model ranks unwatermarked competitor output above freshly minted watermarked positives. A Google-versus-photographs model that never saw a foreign generator accepts 94.6 to 100 percent of every foreign vendor at the photograph median, and scores fal-ai output above Google's own fresh 4K era (AUC 0.946 against it). One outlier cell, OpenAI fresh after JPEG-95 at 0.942 against photographs, carries a storage asymmetry: the positives are stored PNG and the photograph negatives JPEG, and re-encoding both does not equalize that history; it is recorded as an anomaly, not a signal.

What this settles: the learned route at this capacity learns "AI render versus photograph" and nothing vendor-specific. With S6 refuting every hand-crafted feature across the same grid, R0 refuting the phase-locked experts, and M1b bounding what per-image statistics can do against a properly masked mark, three independent directions now agree: a per-image local SynthID detector is not reachable from these pixels by statistical means at any scale we can validate. Provider mark detection requires the provider decoder, which exists as an API for OpenAI and as an access-gated service for Google. The separable local signal that remains is the AI-render versus photograph distinction, whose honest product framing is generation detection, not watermark detection.

2026-08-17: matched-pair manufacturing and the distillation screen

Two measurements moved the "build our own detector" question:

The attack ladder pilot (build_attack_ladder_batch.py + run_attack_ladder_batch.py, batch attack-ladder-pilot-2026-08-17, fifteen arms, every arm submitted exactly once) took three verifier-confirmed OpenAI positives through four attack mechanisms: JPEG q40, 0.5 bilinear resize, sigma-8 noise plus JPEG q85, and 0.35 resize plus JPEG q50. Every single arm returned detected, down to PSNR 20.1 dB and RMS 25.1. The OpenAI mark outlives distortions that destroy the image as a product, and MarkNull-severity attacks (which unwatermark Imagen-3 at PSNR 25.4 in their paper) do not flip this verifier. Matched pairs cannot be manufactured by attacking positives: the flip point sits beyond usable image quality, so same-content label contrast is unreachable on this provider by the attack route. All three re-verified sources stayed detected, so the run doubles as a health check.

The distillation screen (probe_distillation_screen.py) took the other oracle-labeled contrast on disk, the 93 L1 control verdicts (65 detected, 28 not detected; metadata says "no watermark" for all of them, the oracle supersedes it), and asked whether any local score reproduces those verdicts. The shipped folded score reaches AUC 0.613 against a permutation null of 0.485 +/- 0.062, roughly two sigma. The best of 126 candidate scores reaches 0.649 before selection correction. A ridge fitted on the test rows themselves (an explicit upper bound) reaches 0.693, while a ridge fitted on fresh mints against photographs and foreign generators transfers at 0.409, below chance: the mark-specific component does not survive provider-external training. The confound audit shows the two verdict groups matched on median size (1.573 MP both) and mixed across months and generators. Reading: a weak mark-specific trace exists in the natural no-assertion region, with a per-score ceiling around 0.61-0.65 at n=93. Tripling the labeled rows is the cheapest way to either confirm that ceiling or climb it, so the L1 runner resumed on its frozen 315-row plan the same day.

The distillation harness (probe_distillation_train.py, feature cache keyed by content hash over all 329 factory-reachable rows, temporal folds by collection month, nested-penalty ridge, max-statistic null) ran its dry pass at the labels available that evening: 102 hard-contrast rows give temporal-CV AUC 0.632, the 110-row pooled contrast 0.652, and one feature of 124 exceeds the max-statistic null by two sigma (0.724 against 0.657 +/- 0.033). The harness is the acceptance instrument for the growing label set; the decision it will support at n near 280 is whether a weak-but-honest local component ships under an indeterminate-heavy contract or the ceiling is declared.

2026-08-17 evening: decoder substitution and the state of Google access

Access re-checks first, both negative. The Vertex publisher model imagewatermarkdetector@001 returns "not servable in region us-central1" and 404 "not found or your project does not have access" in us-east4, europe-west4 and asia-northeast1, on live credentials; nothing has opened since the 2026-08-15 verification sweep. The hypothesized open release of official image watermark weights ("synthid-suite") does not exist: GitHub, the google-deepmind organization and the Hugging Face registries expose only synthid-text. No official image detector or marker weights are public, so no local distillation of the official encoder is possible either.

Decoder substitution was then tested as a matched-pair factory, on the MarkNull mechanism reading that the mark lives in the decoder's output distribution. build_vae_pair_batches.py froze five oracle-verified OpenAI sources plus their float32 encode/decode round-trips through stabilityai/sd-vae-ft-mse (latent mode, no noise, edge padding to a multiple of eight), PSNR 22.3-32.0 dB, and run_vae_pairs_openai.py submitted the five treated arms exactly once: every arm returned detected. The OpenAI mark is not decoder-bound. It has now survived, with verdicts unchanged, JPEG q40, 0.5 resize, sigma-8 noise plus JPEG q85, 0.35 resize plus JPEG q50 down to PSNR 20.1, and a foreign-VAE reconstruction down to PSNR 22.3. Every same-content pair-factory route is closed on this provider; natural verdicts remain the only label source, which is the L1 extension already running.

The matching Google batch vae-pairs-google-2026-08-17 (three fresh policy-watermarked generations, both arms, six submissions) is frozen for the Gemini web verifier. MarkNull reports its SD-proxy regeneration flipping Gemini verdicts at 25.4 dB, so this batch decides whether a plain reconstruction (no latent optimization) already breaks the Google mark, and with it whether the local negative factory is viable for Google. It requires a browser session and is the next oracle action when one is available.

Resolved 2026-08-18 as vae-pairs-google-v2-2026-08-18, submitted manually by the user through the Gemini web verifier. The first v1 attempt exposed a known trap live: the source arm answered through C2PA Content Credentials (metadata), and the vae arm returned visual analysis with no watermark check; neither is a pixel verdict, and both were discarded. The v2 batch re-encoded every arm through PIL so all arms carry identical chunk sets with the sources' decoded pixels asserted identical, and added two attribution controls. All eight arms returned explicit SynthID pixel verdicts: the three stripped sources detected, the three VAE round-trips detected at 26.6-32.9 dB, the COCO photograph not detected, and the C2PA-signed Adobe Firefly image not detected. Two conclusions. First, the Google mark survives foreign-decoder substitution exactly as the OpenAI mark does, so the plain-reconstruction negative factory is dead on both providers; only MarkNull-style latent optimization remains untested, and it destroys product quality (25.4 dB, SSIM 0.80). Second, the controls give the web pixel verdict its first attribution validation: it stays silent on another generator's AI image, so within this session it read the Google mark or Google pipeline, not generic AI forensics. That makes the earlier comb-vs-broadband and replace-095 verdicts, all answered on Google-generated pixels, retroactively cleaner as mark-presence evidence than they were when recorded.

2026-08-18/19: the three-metadata-dataset program and the shared-component attribution

The negative-factory hypotheses being dead, the working program switched to estimation on positives, organized as three datasets built only from unambiguous metadata: G, Google created rows (4,710, watermarked by documented policy); O, OpenAI watermarked.unbound rows (7,444, all oracle- sampled rows detected); N, images that cannot carry SynthID (C2PA-signed foreign generators plus C2PA-signed smartphone cameras; the 2026-08-18 web control answered not_detected on a Firefly member). The 287 OpenAI no-assertion rows and 70 Google non-created rows stay excluded (L1: 69% of no-assertion rows are actually watermarked), and the 19,792 rows with unknown provenance are not used as negatives.

probe_shared_component.py estimates, per class, the mean origin-aligned opponent-residual pattern over up to six unit-RMS 256-px patches per image, with split-half convergence bounds and photographs as the estimator null (0.002 +/- 0.002, i.e. zero). Results: every AI class carries a strong shared component (Google corpus 0.922 at n=535, OpenAI corpus 0.723, OpenAI fresh 0.978 at n=151, foreign generators 0.597, fresh Google 0.30-0.52 at n=10-11), photographs carry none. Corpus-era components are compression-dominated (50-98% of power on the period-8 comb); fresh-era components are not (4-10% comb).

The attribution of the shared component settles the mark question. In chroma, Google eras correlate pairwise (R-G 0.29-0.59 full-field, 0.25-0.48 after removing the JPEG comb) while OpenAI eras correlate at zero and OpenAI versus Google at zero: the Google component is provider-specific but era-crossing. But it is not the mark. Per-generator decomposition against the fresh 3.1-flash mean shows Adobe Firefly at 0.572 and PixelBin at 0.406 full-band coherence while Microsoft generators sit at 0.03-0.19 and fal-ai at 0.05; band-resolved, Firefly tracks Google in every band including the high-frequency chroma bands where the two fresh Google eras correlate (0.87 in the period-4 render band, 0.30-0.63 above 0.30 cyc/px). The 2026-08-18 oracle control verified Firefly carries no SynthID. Therefore the era-crossing shared chroma component is a renderer-architecture lineage shared by Google's decoders with Firefly and PixelBin, absent from Microsoft, fal-ai and OpenAI, and the era genealogy is visible directly (corpus and 3.1-flash share the period-4 band at 0.92 while 2.5-flash does not, matching the model lineages). No frequency band exists where Google eras correlate and Firefly does not. A one-pixel diagonal shift erases the component completely (0.48 to -0.05), the origin lock already established for the lattice.

This closes the fixed-signature program: every image-independent signature measurable from positives is renderer genealogy, and the mark itself presents no fixed cross-image signature at all, consistent with a content-dependent masked encoder (the M1b ceiling and the DeepMind patent model) and with S4's zero carrier finding for OpenAI. The era-mean patterns are saved as amplified PNGs (shared-component/pattern-*.png) so the pipeline signature is now literally visible. The remaining measurable routes are the OpenAI natural- verdict distillation (L1 at 227/315) and same-content differential analysis inside near-duplicate groups, where a created+edited pair of the same content is the one remaining place a mark-versus-mark contrast could exist; the group miner probe_near_duplicates.py records mixed-assertion groups explicitly for that.

The near-duplicate miner completed over all 12,511 provider rows: 170 groups at Hamming 12, 21 cross-provider, 5 with mixed assertion, largest group 9. Every cross-provider pair is a re-render (all 21 differ in raster size; the closest, 896x1200 versus 895x1200, aligns at only 23.7 dB PSNR), so no same-raster cross-provider pair exists and the cross-provider groups cannot serve as same-content mark contrasts; they remain renderer-attribution controls. The local half of decoder substitution also finished (decoder- substitution.json, 287 photographs and 217 foreign round-trips completing the earlier sets): zero of 124 paired-delta features clear the pre-registered 0.70 acceptance against both negative families (best 0.644), and the OpenAI deltas point the wrong way entirely (0.23-0.30). This agrees with the oracle outcome rather than contradicting it: since the mark survives the round-trip, no mark-destruction signature exists to find, and the null result is the expected shape.

A first product-shaped use of the renderer signature was measured the same day: correlating an image's origin-aligned chroma residual against the frozen era-mean signatures (matched filter over the three Google era patterns) gives AUC 0.914 against foreign generators and 1.000 against OpenAI corpus images, with 0.68 TPR at 10% FPR against foreign. The highest-scoring foreign images are exclusively Adobe Firefly and PixelBin, the two generators that share the Google renderer lineage, so the honest claim is "diffusion-render-family generation detector" (Google, Firefly, PixelBin), not provider attribution, and it inherits the one-pixel origin lock. It is a candidate second pixel signal beside the pipeline lattice, to be frozen under the R0 gate before any product integration.

The local half of the same probe (probe_decoder_substitution.py) runs the round-trip over fresh and corpus Google, corpus and fresh OpenAI, foreign generators and photographs, with a pre-registered acceptance rule: a paired feature delta separating Google from BOTH photographs and foreign generators at AUC 0.70 with a permutation null two sigma below. Separation from photographs alone repeats the render-versus-photo result and is not evidence. Its outcome is in decoder-substitution.json: zero of 124 features clear 0.70 against both negative families (best google_min 0.644).

2026-08-19: aoxo/reverse-synthid scored against the frozen era bank

The Hugging Face dump aoxo/reverse-synthid was downloaded in full to .local-eval/synthid/reverse-synthid-hf (2,197 files, 1,922 images) after the public reverse-SynthID repository had already been audited as a chance-level detector of our lattice, not of the mark. The question left open was whether their solid-color and low-content images were usable as independent copies of the Google-era signatures. probe_aoxo_reverse_synthid.py reuses the shared-component residual, the origin-aligned 256-px unit-RMS patches, and the frozen three-era matched filter. It does not call a provider API.

Their Google-family means are the same renderer genealogy we already have. Chroma R-G of the family mean versus google_corpus is 0.805 on gemini-3.1-flash-image-preview (n=589, split-half 0.958), 0.905 on nano-banana-pro-preview (n=518, split-half 0.938), 0.941 on gemini_random (n=88), and 0.898 on gemini_black_nb_pro (n=91, split-half 0.964). The same means sit at 0.48-0.56 versus google_fresh31 and near zero versus both OpenAI eras and photographs. A two-pixel crop on a gemini_black_nb_pro 2048-px black kills the match (0.408 to -0.063 versus corpus). The amplified family mean is the period-8 vertical lattice. DALL-E 3 is a clean negative for that filter: 0/280 images above the frozen foreign 10% FPR cut (0.122), family-mean chroma 0.105 versus corpus, and the amplified mean is unstructured noise.

Two of their folders are a different shared object and must not be mixed into the Google bank. gemini_black and gemini_white have split-half 0.980-0.983, so they do share something, but chroma versus google_corpus is 0.046 and -0.140. They correlate with each other at 0.527 and with the Google-lineage aoxo families at 0.14 or below. The production folded template accepts 0/20 in each folder.

The production lattice detector follows raster size, not folder label. It accepts 20/20 gemini_black_nb_pro (2048 px, median 0.593) and 20/20 gemini_random (2816x1536, median 0.361), and rejects 20/20 of the leading 1024-px solids in the structured Gemini 3.1 and nano-banana folders. That is the existing large-native scope, not a new carrier.

Adding the four matching aoxo means as extra matched-filter templates makes the generation detector worse on the frozen holdout. Against foreign generators, AUC falls from 0.878 to 0.857 and TPR at 10% FPR from 0.471 to 0.445 (n=535 vs 223). Against the DALL-E bank the same change drops TPR at 10% FPR from 0.991 to 0.907. Foreign and DALL-E scores rise more than Google scores do. The frozen three-era bank is therefore left unchanged. shared-component/era-means-extended.npz keeps the aoxo means as labeled reference copies, including aoxo_dalle3_solid as a negative, and is not an input to the detector. Per-family numbers live in aoxo-reverse-synthid/aoxo-reverse-synthid.json.

Substitution was checked next, because addition can only raise scores. Replacing google_fresh31 with any aoxo Google-family mean does not beat google_corpus alone. On the full in-sample holdout the three-era max is already worse than the corpus mean by itself (full-field AUC 0.878 vs 0.892 against foreign). The leaked in-sample number is only slightly optimistic: ten split-half re-estimates of the corpus mean, scored on the unused half, give AUC 0.880 +/- 0.003 and TPR at 10% FPR 0.480 +/- 0.017. Chroma R-G scoring is the one change that moves the needle on that honest split (0.904 +/- 0.003, TPR 0.597 +/- 0.020). It is a scoring change, not a new signature, and it is not frozen into the detector here.

The dump's actual use is as held-out positives the corpus mean never saw. Against foreign generators, chroma correlation with the frozen corpus mean separates gemini_black_nb_pro at AUC 0.990 / TPR 0.989 (n=91) and gemini_random at 0.935 / 0.750 (n=88). The structured 1024-px solid folders are the same lineage at much lower per-image SNR: nano_banana_solid 0.799 / 0.216 and gemini31_solid 0.801 / 0.153. gemini_black is chance (0.594 / 0.000) and gemini_white is anti-aligned (0.315). Those two folders are near-black and near-white Gemini dumps with a private shared texture (split-half 0.98, pixel std about 3.5), not the period-8 lattice. Numbers: aoxo-reverse-synthid/generation-detector-aoxo-holdout.json and generation-detector-aoxo-substitute.json.

Deflation asks the remaining averaging question: after that lattice is projected out, does a second shared component look like a mark. Projecting a set's own mean out of itself forces split-half near -1 and is discarded as a tautology. The independent case is the one that counts. Removing the frozen corpus mean from the aoxo Google families, or removing the aoxo gemini_black_nb_pro mean from the corpus, leaves a leftover chroma correlation of 0.74 between nano_banana_solid and gemini31_solid. That leftover dies at a one-pixel roll (0.742 to -0.005) and returns at shifts 4/8/16 (0.61 / 0.47 / 0.49). It is uncorrelated with photographs (-0.01), DALL-E solids (0.11), and the OpenAI corpus (0.09). So rank-1 removal peels the JPEG-heavy corpus lattice and uncovers a cleaner origin-locked period of the same renderer family, which is why extra aoxo templates raised foreign scores instead of finding a mark. OpenAI split-half is unchanged by Google-lattice removal (0.723 / 0.978), as required if that component is a different renderer. aoxo-deflate.json and aoxo-deflate-leftover.json.

2026-08-19: remaining local hypotheses

Four locally testable leftovers were closed by probe_remaining_hypotheses.py. None of them is a mark.

Rank deflation. The leftover chroma between held-out nano_banana_solid and gemini31_solid after removing the top K corpus chroma PCs stays at 0.57 from K=2 through K=16 and still dies at a one-pixel roll (0.045). It is not in the corpus JPEG-era subspace. Fitting the PCs on nano_banana_solid itself does remove it from gemini31_solid (0.70 at rank 1, 0.18 at rank 5, 0.12 at rank 8), so the leftover is a low-rank 3.1-family lattice the mixed corpus does not span. Photographs stay at 0.01. Extra ranks do not uncover a shift-robust residue.

Mixed-assertion near-duplicates. All five groups were opened. Three are cross-provider re-renders at different rasters and cannot isolate a mark. The two same-provider OpenAI pairs are same-raster: 1254 px at 29.6 dB (RMS 8.4) and 1054x1492 at 22.1 dB (RMS 20.0). The pixel difference of either pair is uncorrelated with the Google lattice, the OpenAI corpus mean, and the OpenAI fresh mean (best 0.066). Both members of the 29.6 dB pair match the OpenAI corpus fingerprint equally (0.079 vs 0.081). Assertion metadata does not flip a shared residual, which matches the L1 finding that most OpenAI no-assertion rows are still watermarked.

OpenAI shared component. The corpus-era OpenAI mean is origin-locked and periodic: chroma self-correlation 1.0 at shift 0, -0.446 at 1 px, then 0.57-0.68 at 2/4/8/16. Cross-era correlation with fresh gpt-image-1 is ~0 at every shift. The fresh mean itself is origin-locked but not the same period-8 object (self-shift 1 px -0.083, later shifts ~0). Both sit at zero against photographs. This is an era-specific renderer fingerprint of the same class as Google's, not a stable mark.

aoxo gemini_black / gemini_white. Their private shared object is also origin-locked and periodic (black vs white 0.527 at 0 px, -0.502 at 1 px, 0.51-0.53 at 2/4/8/16) and is not the Google corpus lattice (0.046 / -0.140) nor gemini_black_nb_pro (0.139). It is a third phase-locked texture, not a crop-robust mark.

Local averaging, deflation, near-duplicate contrast, and the OpenAI shared component are therefore exhausted as mark routes. What remains needs a provider oracle or a product decision: L1 distillation (OpenAI verifier), H6/H7 (oracle-positive training pairs), freezing chroma scoring under R0 as a generation detector, and H8 as a multi-image product mode. H3 is not a new carrier; a GLRT on the same origin-locked lattice would rediscover S1. Numbers: aoxo-reverse-synthid/remaining-hypotheses.json.

2026-08-19: new search class, lattice death

Hand-crafted templates and cross-image averages are exhausted because they read origin-locked renderer structure. The mark, by publication and by our oracle ladder, survives JPEG-95 and a few-pixel crop; the lattice does not. The next class is therefore: destroy the lattice first, then ask whether any statistic still separates Google from both Firefly and Microsoft.

probe_lattice_death.py applies JPEG-95 and an 8-24 px crop, then scores the folded template, chroma correlation with the corpus mean, and the 124-feature bank. The lattice dies as required: chroma AUC falls to 0.49-0.56 against every family, and the folded template accept rate on Google falls from 0.88 to 0.04. The feature bank does not die. Best-of-124 AUC is 0.796 against Firefly and 0.792 against Microsoft (n=149 vs 78/85). A 200-draw max-stat null on the Firefly comparison has mean 0.600, 95th percentile 0.633, and maximum 0.684; the observed 0.802 never occurs. Size and mean luma are not the confound (AUC 0.51 and 0.35).

The shared winner is feature 106: median B-Y residual autocovariance at lag (2, 3). It is a short-range opponent covariance, not a period-8 comb. On a held-out scoring of that single feature the era gate fails. Corpus Google median -0.015; fresh gemini-3.1 -0.011 (same side); fresh gemini-2.5-flash -0.073 (AUC 0.946 against the corpus, n=10). DALL-E 3 and fresh OpenAI sit with the other generators (0.815 and 0.796). PixelBin, the other shared-lineage renderer, is the nearest foreign (0.663). A mark required by policy on every Gemini era cannot miss 2.5-flash. The leftover is therefore a crop-and-JPEG-stable renderer/era texture, a second pipeline signature beside the lattice, not SynthID.

lattice-death.json and feature-106-holdout.json.

A linear follow-up asks whether the rest of the bank can recover the era that feature 106 missed. Ridge on the 124 JPEG-plus-crop features, trained on 99 corpus Google versus 164 Firefly+Microsoft rows (lambda 100 chosen on an inner hold), reaches AUC 0.852 on the held-out 50 corpus Google images against that foreign pool. It does not transfer: fresh 2.5-flash 0.306, fresh 3.1 0.403, DALL-E 3 0.485, fresh OpenAI 0.418. It does fire on aoxo 3.1 solids (nano 0.962, gemini31 0.978), so what it learned is corpus-era plus low-content 3.1 export texture, not a policy-wide mark. post-lattice-ridge.json.

2026-08-19: L1 finished

The OpenAI control-validation runner completed the 315-row plan. Controls: 203 detected / 285 verdicts (contamination 0.712). Health positives: 27/28 detected, one miss, the same corpus-label error recorded earlier. Two control rows remain error from an earlier unreachable-API window and were not retried, because a request that may have reached the model is not a second submission. The new key was accepted; a 429 after a burst slept 61 minutes and then drained the tail.

On provider-clean labels the frozen patch-local bank still does not separate (patch_local_auc 0.572, shipped OpenAI 0.439, n=283). Distillation on the full L1 set does move: temporal-CV ridge AUC 0.720 on the hard no-assertion contrast (311 rows) and 0.728 pooled, with 8 and 6 univariate features above a 2-sigma max-stat null (best 0.674 vs 0.595 +/- 0.016). In-sample ridge permutation stays at 0.78, so the in-sample fit is still overfit; the 0.72 is the out-of-fold number. That is above the earlier 0.63 at n~110, and still not a production expert. openai-control-validation.json, l1-snapshots.json, distillation-train.json.

2026-08-19/20: OpenAI identity ladder

The 2026-08-17 attack ladder left crop, rotation, flip and grayscale unasked. identity-ladder-2026-08-19 takes the already-verified s1 positive and submits ten frozen rungs once each. All ten returned detected:

Rung	Output	PSNR (resized-back)	Oracle
JPEG q80	1024x1536	37.3	detected
JPEG q20	1024x1536	31.4	detected
crop 2 px top-left	1022x1534	26.0	detected
crop 8 px all sides	1008x1520	21.0	detected
center keep 75%	768x1152	11.8	detected
center keep 50%	512x768	10.3	detected
rotate 90 deg	1536x1024	10.9	detected
rotate 5 deg, expand	1154x1620	13.6	detected
horizontal flip	1024x1536	10.7	detected
grayscale	1024x1536	26.1	detected

Together with the earlier 3x4 ladder (JPEG q40, 0.5x, noise+JPEG q85, 0.35x+JPEG q50, all detected, n=3) the official OpenAI mark, on this seed, survives every identity-class transform that kills the pipeline lattice. The two-pixel crop that zeroes the origin-locked chroma template is not a mark failure. A half-image center crop is not a mark failure. Grayscale is not a mark failure, so the decoder is not chroma-only. 90-degree rotation and a flip are not mark failures, so the decoder is not phase-locked to the original raster axes.

This is n=1 for the new rungs. It is enough to reject several detector hypotheses and to constrain the embedding: a spatially redundant residual, readable from luminance, trained with geometric and JPEG augmentations, not a border, not a fixed tile, not the period-8 lattice. Local detection then has to be a crop-equivariant learned decoder, which is the M1b ceiling, not a better comb statistic. identity-ladder-2026-08-19/manifest.json.

2026-08-20: how the OpenAI mark is applied

The survival envelope plus the paper/patent plus the VAE pairs constrain the embedding method. It is not a new measurement; it is the intersection of the ones already taken.

The encoder is post-hoc and additive in pixel space: after the generator has emitted an image, a separate network writes x' = x + g(x) and a separate network reads it back. That is the paper's definition and the patent's image example. Three oracle facts say the same thing in measurements. Metadata stripping does not change the verdict, so the mark is not a C2PA field. A VAE decoder substitution (sd-vae-ft-mse, 22.3-32.0 dB, 5/5 detected) does not remove it, so the mark is not a latent of that decoder. Replacing the generator would not remove the capability, which is why Google and OpenAI can share a SynthID family name and still fail each other's official decoder.

Where it is written. Grayscale (PIL L, then RGB) stayed detected, so the decoder reads luminance: stacking L,L,L zeroes chroma, and a chroma-only carrier would have gone silent. JPEG q20 stayed detected, so the carrier is not the high-frequency DCT tail that classical DWT-DCT uses. A center crop keeping half the area stayed detected, so the residual is spatially redundant rather than a border or a single tile. A two-pixel crop stayed detected while our origin-locked lattice dies, so the written residual is not that lattice. 90-degree rotation and a horizontal flip stayed detected, so g is not a phase-locked overlay on the original raster axes.

How g is shaped. M1b at matched RMS put a fixed mid-band mark at AUC 0.670, a keyed mid-band mark at 0.669, and a keyed perceptually masked mark at 0.546. Masking, not keying, is what hides it from hand statistics: amplitude tracks local texture, local SNR is flat, and there is no quiet region to average. The patent's U-Net with optional key injection is the architecture that produces exactly that residual. Content dependence (g(x) rather than a fixed g) is why averaging thousands of OpenAI images never yielded a mark template, only an era-specific renderer fingerprint.

The working model of application is therefore: a provider-specific neural encoder, run after generation, adds a perceptually masked luminance residual in the mid-band, redundant across the frame, trained to survive JPEG, crop, resize, flip and small/90-degree rotation. The matching decoder holds the provider key as weights. Nothing in the local pixel statistics we can write by hand is that decoder.

2026-08-20: census of embedding methods

Every plausible way to write a SynthID-class mark, sorted by what the measurements already forbid. Identity-ladder rungs (crop, rotation, flip, grayscale, JPEG q20) are n=1 on OpenAI s1. Attack-ladder and VAE pairs are n=3 and n=5. Google VAE is n=3. A method that survives those tests is not proved; a method that dies on them is closed.

Ruled out

Method	Why it is closed
C2PA, EXIF, filename, chunks, or any sidecar	Official OpenAI path strips AI metadata, proves pixels unchanged, still detected
LSB / bit-plane steganography	JPEG q20 and sd-vae-ft-mse both destroy LSBs; both left the mark detected
Classical DWT-DCT / imwatermark bins	That family dies near JPEG q90; SynthID stayed detected at q20 and at 0.35x+q50
Chroma-only residual	Grayscale zeroes R-G and B-Y; oracle still detected
High-frequency-only tail	JPEG q20 and VAE keep mid-band and discard the tail; mark remained
Border, corner, or single-tile stamp	Center keep-50% still detected
Origin-locked periodic lattice (our S4/S1 object)	2 px crop kills that lattice and not the oracle; Firefly shares the lattice and has no SynthID
Axis-locked overlay on the original raster	90 deg rotation and horizontal flip still detected
Mark stored only in one generator's VAE latent	Decoder substitution through sd-vae-ft-mse left 5/5 OpenAI and 3/3 Google detected
In-model watermark baked into diffusion weights	Paper: post-hoc f: X -> X on an already generated image; swapping the generator does not remove the capability
One shared decoder for Google and OpenAI	OpenAI verifier returned not_detected on a Google positive and detected on OpenAI positives
Fixed additive template shared across images	OpenAI corpus vs fresh split-half cross-era ~0; averaging recovered renderer genealogy, not a mark
Keying alone as the reason hand features fail	M1b: fixed vs keyed mid-band 0.670 vs 0.669; masked keyed 0.546

Still possible, not forced

Method	What would still fit
Additive x' = x + g(x) vs a learned invertible map	Additive is the patent example; a small invertible network is not ruled out if it is JPEG-and-crop stable
Luma-only vs luma-plus-chroma	Luma is required; extra chroma is allowed because grayscale kept the luma copy
Mid-band spatial residual vs mid-band learned latent of the watermarker itself	VAE test only kills that VAE; the encoder may still think in its own latent
U-Net vs other conv/attention generators	Patent describes U-Net; any local-and-global residual net that masks to texture would survive the same tests
Resize to a trained size, mark, resize back	Patent alternative; compatible with scale survival; decoder would canonicalize or be scale-trained
Ensemble of encoder/decoder pairs per provider or epoch	Patent continuation; explains Google vs OpenAI silence and era-specific keys without changing the method class
Per-image payload / secret injected into g	Presence detection does not need payload recovery; optional in the patent
Training augmentations beyond the ones we hit	Elastic warp, extreme downscale, JPEG q5, 10% area crop are untested; they may be outside the trained set
Slightly different Google vs OpenAI implementations of the same class	Same method, different weights, possibly different trained-size or mask strength

Forced by the intersection

Property	Forced by
Written in decoded pixels, after generation	Metadata strip; paper definition
Readable from luminance	Grayscale detected
Spatially redundant across the frame	Keep-50% detected
Not phase-locked to the source origin or axes	2 px crop, rot90, flip detected
Mid-band, JPEG-stable	q20 and 0.35x+q50 detected
Provider-specific decoder weights	Cross-provider not_detected
Content-dependent residual, perceptually masked	No shared template under averaging; M1b mask gap

The method class is therefore one: a provider-keyed neural residual in luma mid-band, post-hoc, redundant, geometry-trained. Variants inside that class (U-Net vs other nets, optional payload, resize-to-canonical, ensemble of pairs) remain open and do not change the detector architecture. Everything outside that class is closed.

2026-08-20: OpenAI kill ladder, first not_detected

Identity and attack ladders never silenced the official decoder. The kill ladder on the same s1 seed asks which frozen destruction first does.

Rung	Output	PSNR	Oracle
JPEG q10	1024x1536	29.1	detected
JPEG q5	1024x1536	26.3	detected
center keep 25%	256x384	9.9	not_detected
center keep 10%	102x153	10.5	not_detected
Gaussian blur sigma 3	1024x1536	26.9	detected
Gaussian blur sigma 8	1024x1536	23.3	not_detected
scale 0.20	204x307	27.0	detected
median 7	1024x1536	28.6	detected
elastic alpha 12	1024x1536	40.7	detected
posterize 4 bits	1024x1536	29.5	detected

The flip is not pixel count. Scale 0.20 has fewer pixels than keep-25% and stayed detected, because it still holds the whole frame. Keep-50% was detected on the identity ladder; keep-25% is the first crop miss. Blur sigma 8 is the first full-frame miss; JPEG q5 is not. Mild elastic (40.7 dB) is not reverse-SynthID's 18-24 dB warp and did not flip.

So the official decoder needs a large fraction of the original scene and is more sensitive to spatial low-pass than to DCT quantization. That is not a remover: blur-8 at 23 dB and a quarter-frame crop destroy the picture. It is the first oracle-negative envelope on OpenAI. n=1 seed. kill-ladder-2026-08-20/manifest.json.

The in-between batch flip-ladder-2026-08-20 on the same seed:

Rung	Output	PSNR	Oracle
blur sigma 5	1024x1536	25.1	detected
blur sigma 6	1024x1536	24.4	detected
blur sigma 7	1024x1536	23.8	not_detected
center keep 40% linear	409x614	10.5	not_detected
center keep 33% linear	337x506	10.1	not_detected
elastic ~24.6 dB	1024x1536	24.6	detected
elastic ~20.8 dB	1024x1536	20.8	not_detected

Flip points, still n=1: Gaussian blur between 6 and 7 (24.4 dB vs 23.8 dB); crop between keep-50% linear (detected on the identity ladder, 512x768) and keep-40% (409x614); elastic between 24.6 dB and 20.8 dB. JPEG q5 remains detected. No rung that silences the decoder is a usable picture. flip-ladder-2026-08-20/manifest.json.

2026-08-20: OpenAI add ladder

The crop miss could have been "too few marked pixels in the upload" or "too little of the original scene". Adding unmarked pixels distinguishes them. Same s1 seed, donor COCO 000000000139.

Rung	Upload	Oracle
pad white, original 70% linear	1463x2194	detected
pad white, original 40% linear	2560x3840	detected
inset on photo, 70% of 1600 canvas	1600x1600	not_detected
inset on photo, 40% of 1600 canvas	1600x1600	not_detected
blend 50% with photo	1024x1536	detected
blend 25% marked / 75% photo	1024x1536	detected
horizontal concat with photo	2048x1536	not_detected
Gaussian noise sigma 16	1024x1536	detected

White padding to 40% linear stays detected, while cropping the same seed to 40% linear was not_detected. The decoder is not counting marked pixels in the upload; it is reading coverage of the original scene. Uniform pad is ignored. A second unmarked photo beside the intact marked half (hstack) silences it, as does pasting a resized copy onto a photo (inset, with an aspect-ratio confound). Mixing unmarked pixels into every location at 75% still detected, so amplitude 0.25 of the residual is enough when it is present everywhere. n=1. add-ladder-2026-08-20/manifest.json.

All OpenAI oracle ladder numbers are copied into tracked docs/synthid-removal-research.md so they survive the gitignored eval directory.

Replication on s2 and s3 of the s1 flip points is complete. Blur sigma 6 vs 7 and crop keep 0.50 vs 0.40 hold on 3/3 seeds. White pad 40% stays detected and hstack stays not_detected on 3/3. Blend 25% stays detected on 3/3. Native-aspect inset onto a 2x photo canvas is not_detected on 3/3 (s1 added). Elastic at ~21 dB does not generalize: s3 remains detected at 22.0 dB while s1 and s2 miss. The durable crop and blur flips are therefore the ones to keep; elastic PSNR is not a stable kill threshold.

Surround ladder on s1: photo fill at the white-pad geometries is not_detected at both 70% and 40% linear, so extra canvas only kills when the pad is a photograph. A side strip of unmarked photo stays detected at 10% and 25% of width and becomes not_detected at 35% and 40%. The decoder treats a large unmarked scene in the same upload as a whole-frame negative, while uniform white is ignored. Numbers in docs/synthid-removal-research.md.

Photo-pad vs white-pad holds on 3/3. The hstack unmarked 25% vs 35% split holds on 3/3. A photo frame around s1 stays detected at 5/8/10/12% each side and becomes not_detected at 15%. Frame 15% holds on s2 and s3. Frame 12% does not: s2 is already not_detected.

2026-08-20: close the local OpenAI mark hunt

Decision: stop searching for a local OpenAI SynthID detector or a quality-preserving pixel remover on this evidence. The official decoder's flip points are now replicated (blur 6 vs 7, crop keep 0.50 vs 0.40, photo surround, hstack 25 vs 35, JPEG q5 still detected, all 3/3 except elastic PSNR and frame 12%). Every not_detected that is not a collage is below usable quality. Distillation on L1 labels remains 0.72 temporal-CV against a 0.78 in-sample permutation, so it is not an expert.

What the product already ships is the honest remainder: signed provenance for the watermark, and detect-synthid / identify's pipeline_lattice for the origin-locked generation lattice, labeled as not SynthID. The unshipped chroma matched-filter (AUC about 0.90 vs foreign on the frozen holdout) is a second generation-family candidate and may only enter runtime under the R0 gate, never as a watermark claim.

Oracle numbers: docs/synthid-detector-research.md and docs/synthid-removal-research.md.

Open questions that are not a repeat of closed hunts, now that the official OpenAI decoder can be queried:

1. Self-concat. hstack of marked|unmarked is not_detected on 3/3. hstack of marked|marked (same seed, two copies) is untested. If that stays detected, the kill is a second scene, not extra canvas. If it misses, any two-panel layout drops the score.
2. Surrogate. The ladders plus L1 are a small labeled set of (image, transform, official verdict). A student that copies those verdicts would let attack search run offline. It is not a mark decoder until it beats the 0.72 L1 temporal-CV and the 0.78 in-sample permutation on held-out seeds.
3. Luma-only blur between sigma 6 and 7 at PSNR above 30 dB. Closed 2026-08-22 as a quiet-remover class together with Bayer, VNG, barrel, and scanline: none flipped the OpenAI oracle below scramble cost. The only full-frame residual kill remains ~24 dB.
4. The chroma matched-filter remains a generation-family candidate under R0, never a watermark claim.

Item 1 is the cheapest untested discriminator. Item 2 is the only path to a local stand-in for the official decoder. Item 3 is the last quality-removal check worth an oracle burst.

2026-08-20: cardinal directions, including the text analog

SynthID-Text is not a recipe for the image encoder. Text is biased at sampling time among near-synonyms; the published image system is a post-hoc map f: X -> X. The useful borrow is the detector shape: a keyed weak statistic pooled over many tokens, so you need a long enough span and unmarked tokens vote against. That is compatible with what the OpenAI image oracle already did (keep-0.40 misses, whole-frame 0.20x hits, white pad ignored, photo surround votes no). It does not revive averaging or a fixed tile.

Directions that are actually untested, not a new comb:

• Token analog. Submit the four quadrants of a known-positive separately. 4/4 detected means a locally redundant residual. 0/4 detected while the full frame hits means a global pool. Mixed means content-dependent support. Cheap, four uploads, not a transform we have run.
• Layout scramble. Cut the same positive into a grid of tiles, permute, reassemble, upload. Local-and-pool marks can survive; layout-locked or object-aligned g(x) should drop. Orthogonal to blur and crop.
• marked|marked concat. Distinguishes "second scene" from "any two-panel layout".
• Self-keyed residual. Text hashes the prefix; an image analog hashes a thumbnail and chips the residual. Averaging across images then must fail, which it did. The untested check is whether an image's high-pass residual is a stable function of its own low-pass thumbnail across seeds. A yes would be a local handle without the provider key. A no leaves only the official decoder.
• Surrogate of the oracle, not of the mark. Train on the labeled ladder plus L1. Risk: it learns collage/blur/crop, not SynthID. Gate it on held-out seeds and on JPEG q5 still calling detected.

Tree-Ring-style latent Fourier marks are already a poor fit: VAE substitution left the OpenAI mark up. Do not reopen band templates, aoxo, or elastic PSNR as a threshold.

Token-ladder on s1, 2026-08-20: all four quadrants detected at 512x768. The mark is not a single global code that needs the whole frame; a corner half is enough, matching keep-0.50. hstack and vstack of two copies of the same marked image are both not_detected. Photo surround was therefore misread: the decoder also fails when both halves are marked. A two-panel layout is the kill, not unmarked pixels. 4x4 scramble stays detected; 8x8 scramble misses, so the local support is larger than ~128 px tiles and can live in ~256 px tiles. Together with scale-0.20 still detecting the full scene, the official decoder looks like a pool over photograph-shaped regions of sufficient span, plus a gate that rejects obvious collages.

An independent Claude Code pass (2026-08-20, tools disabled) argued the collage gate may be an artifact of canonical-canvas resampling: the decoder resizes the whole upload, so a 0.20x full scene is restored, while a photo-pad or two-panel layout leaves the mark at half canonical scale. That one mechanism would reclassify most collage not_detected as preprocessing, leaving blur-7 as the only confirmed residual kill. Cheapest falsifier: submit 0.20x as-is (known detected) versus the same 0.20x pixels centered on a white canvas of the original size. Identical marked pixels, only canvas scale changes. If the padded copy misses, scale-on-canvas is real. If it still detects, the collage/photo-content gate stands. A second independent pass the same day used Claude Code Opus with full tools and Codex gpt-5.6-sol at high reasoning. Full texts: .local-eval/synthid/prc-oklab-attack-2026-08-15/agent-claude-plan.md and agent-codex-plan.md.

Merged ordered plan from those two independent reviews, OpenAI first:

0. Repair L1 evaluation (0 calls). Codex: the 0.72 is leave-one-date-out on ~4-image folds, includes health positives, and the 0.78 permutation is in-sample. Forward-temporal control-only reanalysis is about 0.64-0.65. Use that as the baseline, not 0.72 vs 0.78.
1. The 0.20x-on-native-white discriminator was run anyway on s1. The 0.20x file alone was detected; the same pixels on a 1024x1536 white canvas are not_detected. Stretch 2x width is detected. Centre-crop of marked|marked back to native size is detected. Pure aspect and pure duplication at native size are not the hstack kill. Scale of the marked region on the uploaded canvas is real. It does not replace pad_photo vs pad_white, which still shows unmarked photographic content as a second mechanism.
2. L1 repair: control-only forward-temporal nested ridge is 0.649 / 0.641, not 0.72. June-15 cut beats a 40-run nested permutation at p=0; June-30 cut p=0.075.
3. E3 leave-one-encoder-out on 80 COCO images: ridge and a 4-layer conv both fail to transfer to TrustMark (AUC 0.51 / 0.51). Ridge transfers only to dwtDct at 0.675. After VAE laundry of both classes: all three held-out families are chance or worse (0.41 / 0.49 / 0.47). The 0.65 gate fails. A local blind detector of this mark class is not learnable here at this capacity and data.

Independent debate 2026-08-20 (Opus full tools vs Codex gpt-5.6-sol high): Codex votes CLOSED for the whole local-blind program. Opus splits: blind transfer CLOSED, Google CLOSED, OpenAI OPEN by one experiment (E4: graded phase rotation in 16-32 px, stay open only if some theta < pi kills the oracle while a keyless paired AUC on that perturbation stays below 0.75). Both texts: agent-debate-claude.md, agent-debate-codex.md under the eval directory. Third-voice merge: honor E3 as closing surrogate-encoder transfer; do not reopen templates. The only remaining OpenAI experiment that is not shopping is a short E4: three seeds times theta in {pi/4, pi/8} in the 16-32 annulus, with a matched 64-128 control at the same theta. If nothing below full scramble flips 2/3 seeds, close. If it flips, the pair layer is real and D4 may be attempted once. That is 12 calls, not 57.

Unbounded open pass 2026-08-20 (Opus and Codex gpt-5.6-sol high, full tools, local measurements, no oracle). Texts: agent-open-claude.md, agent-open-codex.md.

Opus measured TrustMark mark/content power in the carrier octave at -33.9 dB and showed a keyless energy detector's AUC is theoretically ~0.51 at s1 size and ~0.54 at 4K; campaign keyless numbers (0.53-0.57) sit on that ceiling. E4 pair-by-band-destruction is then dominated by the attack (+24 dB over the mark at the gentlest killing theta). Micro-warp desync failed. An external HF "gpt-image-synthid-detector" scores 0.50 vs COCO (positive on 80% of photos) but 0.69 on L1 official labels; the 0.70 plateau is real and unnamed.

Codex measured that an 11-number geometry vector (size, aspect, modulo 64) gets forward-temporal AUC 0.79-0.80 on the same 283 L1 controls where pixel ridge is 0.64. Raster geometry currently beats every watermark representation as a predictor of the official verdict. TrustMark-P analog: survives 0.20x 12/12, dies on crop and scramble; its decoder center-crops to 224, so lethal octaves can be decoder coordinates.

Merged next action if the local hunt continues at all: zero-API, geometry-conditioned L1 (Codex) plus naming the 0.70 (Opus). Do not run E4. Do not train a CNN on raw L1 bits.

Ran 2026-08-20. ChatGPT rows are all not_detected (21/21). After dropping them, geometry still beats pixels (0.72/0.78 vs 0.63/0.67). Largest mixed API size 1254x1254 is chance (p=0.45). Self-keyed 16-32 map is 0.41 vs L1-neg and 0.50 vs COCO. L1 official bits remain unidentifiable as a mark task on this corpus. Do not train a CNN on them.

Detector debate 2026-08-21 after gpt-image-2 pairs (Opus, Codex gpt-5.6-sol high, this session). Notes: agent-detector-claude.md, agent-detector-codex.md. Opus: photo matched-filter ceiling from the flat 16-32 amplitude (0.11 RMS) vs in-band photo content (12.2 RMS) is d'0.93 / AUC0.75 even with a perfect template; P5 needs d'>=4.37, a 13.4 dB deficit unless the photo-domain mark is louder than the flat measurement. Next: 16-32 jamming titration on the fish edit and the gray-128 flat (~14 oracle calls) to measure that amplitude. Codex: do not reopen keyless search; train a frozen wavelet CNN only on procedural flat-to-texture pairs, lock >=128 photos, require paired AUC LCB>0.65 plus a five-file causal panel. Merge: titration first. A student cannot beat a matched filter if the amplitude is as quiet as the flats; minting 128 photos before that number is architecture shopping. Retract L1 matched-filter AUC 0.701 as size confound.

Titration ran 2026-08-21. Gray arm is not a coding-gain calibrator (any extra energy, including 4-8 px, is not_detected). Fish 16-32 additive jam stays detected through a=16 / 21.9 dB and flips at a=24 / 18.6 dB; 4-8 px at a=16 stays detected. 4-8 px jam at a=24 / 18.6 dB stays detected, so the 16-32 flip is band-specific. Phase scramble of 16-32 remains the cheaper kill at 24.3 dB. The close a*_photo < 0.53 failed. Do not mint a 128-photo student on that number. Do not treat additive in-band noise as a photo remover.

2026-08-21: source-matched pairs and titration

Full numeric log: docs/synthid-detector-research.md and docs/synthid-removal-research.md.

gpt-image-2 can mint encoder-on/off pairs. Flats are clean (PSNR 38-49 dB, oracle not_detected to detected) and share a highpass residual (LOO 0.645) that does not transfer to photographs. The 16-32 carrier is causal on both flats (scramble 51 dB not_detected) and the fish photo (scramble 24.3 dB not_detected; pyramid wipe 48 dB still detected; additive jam flips only at 18.6 dB, while equal-energy 4-8 px jam stays detected).

That does not open a local photo detector: the flat stamp is not the photo mark, L1 bits remain a geometry/export task, and the photo jam amplitude never fell below the preregistered 0.53 RMS close. Product remainder unchanged.

Future analog track, not a current experiment: a self-owned encoder in the 16-32 px midband (thumbnail-keyed phase, JND mask) as a research stand, never as a Gemini/OpenAI detector. Future evasion track, only with oracle quota: a graded phase shift in 16-32 at PSNR > 30 dB on s1-s3. Collage, photo-pad, and small-on-canvas stay presentation bypasses, not removers.

2026-08-22: Model 1 result, provider CLIP, lattice re-check, Bayer close

Numeric log: docs/synthid-classifiers.md and docs/synthid-removal-research.md.

Model 1 is the AI-versus-camera research result, not a watermark detector. Finetuned CLIP-L, 224 letterbox, last two vision blocks: Kodak 0/24, locked Open Images fresh FPR 1.7%, AI-test TPR 93.0% (OpenAI 93.2, Gemini 90.5, FLUX hold 92.7). False positives on Open Images are graphics, CGI, and cutouts, not Gemini. It does not enter identify. CLIP-H and DINOv2-giant are not the result.

Provider CLIP (openai / google / other) and an OpenAI-plus-Gemini union fail Firefly. The 124-d residual bank remains the Google-lineage lattice, not an AI-or-not feature set.

Production pipeline_lattice re-check (n=628): Google 56%, Firefly 18%, PixelBin 14%, OpenAI 1%, photographs 0. A 2 px crop kills the positives. Name stays generation-pipeline lattice.

Bayer bilinear, Bayer VNG, upscale-then-Bayer, barrel, and scanline left OpenAI SynthID detected on s1 and s2. 16-32 phase scramble still flipped both. VNG on s2 is dirtier than scramble and the mark remains. Camera pipeline and mild geometry are closed as quiet removers. s3 and the fish photo were not submitted in this pass (OpenAI 429). 4. E2 per-octave phase scramble on s1, replicated p16_32 on s2 and s3: destroying periods 16-32 px is not_detected 3/3; 2-4, 4-8, 8-16, 32-64, 64-128 stay detected. The carrier octave is 16-32 px, not the lattice band. Keyless energy and RGB phase-agreement in that octave on L1 are AUC 0.53 and 0.57. 5. Google gets no independent hunt. If (3) transfers, validate on a small Gemini-app batch against Firefly/PixelBin. If (3) fails, say so.

Working decoder hypothesis, not proved: a net trained on single whole photographs, probably after a canonical resize. That one story fits JPEG q5 (still one photo), scale 0.20 (same scene), white letterbox (still one photo), keep-0.50 vs 0.40 (enough vs too-tight crop of the scene), photo-pad and any two-panel concat including marked|marked (not a single photo), scramble 4x4 (still reads as one image) vs 8x8 (a mosaic). It is a hypothesis about the decoder, not a new encoder. Falsify it with a concat that still looks like one photograph, or a mosaic that still looks like one.

Ontology split, forced by marked|marked missing while every quadrant of the same seed detects: an official not_detected is not one event. Collage, photo-pad, two-panel self-concat, and 8x8 scramble can be a presentation gate (this upload is not a single photograph). Blur sigma 7, keep-0.40, and JPEG-surviving full-frame views sit on the single-photo manifold and are the only current candidates for actually damaging the residual. Crop-keep-0.40 may be either: too little scene for the residual net, or a zoom the preprocessor treats as a different image. Do not train a surrogate on mixed collage misses if the goal is a mark student; those labels are the gate.

Production hardening roadmap

The next milestone is a precision-first production detector system, not one unqualified binary classifier. "Bulletproof" means that every positive verdict is tied to a frozen, independently challenged expert; every negative verdict is limited to a sensitivity-certified scope; and every other case is an explicit abstention. It does not mean claiming that an undocumented watermark can be recognized after every possible edit or future encoder change.

Current starting point

Route	Evidence already obtained	Production gap
Gemini fixed native, 1-10 MP	Frozen period-16 template; 0/5,000 observed-geometry COCO controls and 0/5,000 generated-geometry controls	Too few matching-oracle positives to estimate recall by geometry, epoch, and product surface
Gemini large native, 10-18 MP	7/7 official pixel positives; 0/2,637 feature-unseen controls; 0/41 post-freeze Open Images controls	Positive sample is too small; JPEG-95 and JPEG-90 each reduce 7/7 to 0/7; the fresh negative acquisition is incomplete
Gemini registered cascade, 0.25-10 MP	Registered-v3 plus bounded opponent fallback; the fallback recovered 49/49 lossless 0.5x-0.75x views from seven official positives, rejected 1,790 measured period-8 codec crossings, and retained the earlier 0/1,000 post-freeze control result	Provenance labels are not enough for a provider-wide recall claim; lossy retranscodes are rejected; additional encoder epochs remain abstentions
OpenAI official remote	Metadata-free, pixel-identical web smoke: 2/2 OpenAI positives detected, while one Google positive and one COCO control returned not detected; one preselected OpenAI-positive PNG also returned SynthID detected through the credentialed SDK path; mocked transport and parser tests pass	JPEG, WebP, approved negative transport cases, operational limits, privacy behavior, and live error paths are not yet certified
OpenAI local	Period-8, DTCWT, codeword, external-decoder, and generic-watermark transfer branches all failed their locked or fresh checks	No local route may ship without authorized pixel-oracle labels or source-matched counterfactual pairs

These rows are separate claims. The OpenAI official verifier must never be used as a fallback label for Gemini, and the local Gemini carrier must not be reported as OpenAI or even as Google-exclusive until provider exclusivity is measured against other SynthID deployments.

P0: freeze the product contract before collecting more scores

Introduce one provider-neutral result envelope around the existing backends:

• status: detected, not_detected, unsupported, indeterminate, or error;
• signal_family, provider_scope, backend, detector_version, and calibration_version;
• decoded geometry, input format, score, threshold, and the exact support reason;
• pixels_preserved and metadata_used_for_verdict audit fields;
• a calibration-manifest digest and an optional decoded-pixel audit digest that stays in the local qualification manifest and is never emitted in normal user output or centralized telemetry;
• explicit caveats for lossy input, spatial resampling, unknown encoder epoch, and remote upload.

At the product boundary, a miss from the current positive-only Gemini experts maps to indeterminate, not to a clean-image claim. not_detected is reserved for an official provider verdict or a future local scope whose sensitivity gate has passed. unsupported, remote failure, malformed response, quota exhaustion, and verifier refusal remain distinct states and never collapse into a negative.

The expert registry owns support predicates and precedence. Fixed, registered, large, and future codec experts keep separate identities and calibration artifacts. A router may select one predeclared expert from input scope, but may not OR together overlapping scores or relabel the default expert after seeing its value. Offline commands never invoke a remote verifier implicitly.

The first P0 runtime slice is now implemented. Local results return detected, indeterminate, or unsupported, include an explicit support or miss reason, and expose signal_family, provider_scope, backend, metadata_used_for_verdict, and pixels_preserved. The official OpenAI result uses the same audit fields while retaining the provider's documented detected/not_detected outcomes. Calibration-manifest digests and a unified error envelope remain qualification work rather than being synthesized from the historical research files.

P1: build the immutable qualification corpus

All qualification media stay outside the public repository. The tracked artifacts are schemas, hashes, split manifests, aggregate reports, and cleared synthetic fixtures.

For Gemini, collect matching-oracle labels before any new threshold fitting:

• at least 200 decoded-pixel-unique native positives for the fixed route, balanced across five pixel-count bins, landscape, portrait, and square geometry, at least three collection windows, and every available Gemini image surface;
• at least 100 official native positives for the large route, balanced across 10-12, 12-15, and 15-18 MP, the three aspect families, and at least three collection windows;
• at least 200 source-disjoint transformed positives before reconsidering the registered route, with the untouched parent verified first and every derivative grouped with that parent;
• at least 10,000 freshly frozen decoded-pixel-unique negatives per claimed local route. Each route needs cameras, scans, screenshots, flat graphics, synthetic textures, other generators, other watermark families, resampling aliases, and same-provider oracle-negative examples. No source family may dominate the aggregate result.

Complete the pending 3,000-image post-freeze Open Images large challenge rather than treating the current 41 completed files as a final estimate. Deduplicate every item against all prior train, development, threshold, and challenge corpora by decoded hash and perceptual group.

For OpenAI, prepare a preregistered transport-qualification set with previously established ordinary-use verdicts across PNG, JPEG, WebP, portrait, landscape, square, alpha, maximum practical dimensions, and multiple model/date windows. Run it through the SDK only within the published usage policy and the account's authority. If the required batch volume or retention terms cannot be approved, the remote backend remains opt-in preview rather than being called fully qualified. These cases validate transport and response semantics; they are not training labels for a local detector.

Every split is frozen by parent group before scoring. Mutation tests must prove that swapped providers, changed bytes, duplicate pixels, derivative leakage, indeterminate-as-negative conversion, and a wrong expert identity invalidate the run.

P2: finish the Gemini local expert bank

Rebaseline the fixed and large routes only on oracle-confirmed positives and fresh negatives. Do not tune the existing holdouts again. Each failed gate creates a new detector version and a new untouched test partition rather than a patch to the same test set.

The all-resolution program is a registered bank of measured carrier periods, not an assumption that arbitrary dimensions imply arbitrary scale. Test these branches in order:

1. native period-16 fixed and large experts, with no threshold changes unless the new locked corpus rejects them;
2. integer-period experts for periods 8, 10, 12, 14, 18, 20, and 24, each with its own search-corrected null calibration;
3. a fractional-period expert that estimates period on one image region and scores a disjoint region, preventing the same noise peak from selecting and proving itself;
4. codec-specific experts trained and evaluated symmetrically on JPEG and WebP pixels, never by requiring a native file to survive a second lossy encoding;
5. crop and screenshot experts only if phase registration and boundary search pass a fresh multiple-search challenge.

Candidate features remain restricted to watermark-mechanism evidence: registered residual tiles, signed opponent-color phase, cross-region phase agreement, circular phase statistics, and directional wavelet residuals. Scene semantics, thumbnails, filenames, container size, metadata, and provider logos are forbidden. A learned model may be introduced only after a classical expert passes the same source-matched gate, and it must consume high-pass residuals rather than RGB content.

For each new expert, fit on train, choose features on validation, calibrate on a separate null partition, and evaluate once on locked and later temporal tests. Report per-parent paired differences and confirm claimed improvements with a sign test. A larger aggregate AUC does not compensate for a failed hard-negative stratum.

Ranked detector hypothesis backlog

The evidence gap blocks threshold promotion, not mechanism research. The following hypotheses are materially different from the already rejected extra color spaces, raw FFT magnitude, standalone SWT/DTCWT, multifractal summaries, folded variance, spatial lag products, and provider-versus-control CNNs. They are ranked by expected information gain and must still follow the P1 split discipline.

Execution state as of the 2026-08-16 audit (the entries below keep their original wording as the research designs; this block records what happened):

#	State	Outcome
H1	Executed 2026-08-14	Split-confirm lattice synchronization works: 20/20 transformed positive views at exact expected periods, 0/800 locked negative views. But the lattice it locks onto is the origin-anchored generation-pipeline lattice refuted as the watermark under S1/R0, and the broad non-Google challenge accepted 3/276. Closed as a watermark-detection route; remains the mechanism baseline
H2	Rejected 2026-08-14	The complex spectral-correlation statistic did not beat its own phase-preserving mutation control; rejected without consuming another locked scale partition
H3	Partially executed	Split-confirm selection and the content-matched same-image null ran inside the H1/H4/H5 probes; the full GLRT with empirical p-values folded into the conformal cascade is still open
H4	Rejected as a standalone detector 2026-08-14	Content-whitened matched filter added no boundary over the phase-preserving spatial template; survives only as an H1 synchronization refinement, calibrated inside that branch
H5	Advanced as research confirmation	The unknown-payload patch-agreement statistic retained positive margins under JPEG and crop; explicitly not a runtime threshold, and it inherits the S1 scope question because it reads the same periodic structure
H6	Blocked	Requires the matching-oracle positive corpus (P1 prerequisites, P3 boundary for OpenAI); not started
H7	Blocked	Same corpus dependency as H6; not started
H8	Not started	A separate batch product mode, not a substitute for the single-image detector

H1: recover the complete reciprocal lattice

The current scale registration estimates mainly one carrier period. A resize, rotation, shear, screenshot, or mild perspective warp transforms a two- dimensional reciprocal lattice, not one scalar period. Fit two lattice basis vectors from harmonic peak pairs with a robust consensus estimator, then demodulate or resample the residual into canonical carrier coordinates.

Estimate the lattice on one fixed checkerboard of spatial blocks and score the held-out blocks. This prevents content peaks used for synchronization from also proving detection. Compare translation-only, isotropic-scale, affine, and mild projective models in that order, charging every model and harmonic candidate to the same maximum-statistic calibration.

Gate: improve paired sensitivity on resized, rotated, and screenshot positives without raising the frozen hard-negative upper bound. If affine registration selects natural resize lattices as often as watermark lattices, reject the branch rather than adding a semantic veto.

H2: complex cross-spectral cyclostationary coherence

The rejected OpenAI cyclostationary screens measured folded magnitude, periodic variance, and spatial lag products. They did not test the complex spectral correlation density F(k + alpha/2) * conjugate(F(k - alpha/2)) across disjoint spatial blocks. A spread-spectrum periodic modulation can couple neighboring frequencies even when payload signs cancel the cross-image mean.

Measure normalized complex coherence only at preregistered cyclic frequencies and compare it with same-radius off-carrier frequency pairs. Preserve phase and cross-channel covariance instead of averaging magnitudes. Fit frequencies on train, choose one statistic on validation, and confirm on locked images and fresh same-provider controls.

Gate: the carrier-to-off-carrier contrast must remain positive within images, across temporal windows, and after symmetric codec transforms. A difference that appears only between providers is a renderer statistic and is rejected.

H3: select-confirm GLRT with a same-image null

Large external control sets remain necessary, but each image can also supply a content-matched null. Estimate period, phase, orientation, and channel direction on one region set. On disjoint regions, compare the selected carrier against neighboring off-lattice bins, orthogonal orientations, phase-scrambled templates, and nonselected lattice hypotheses. Convert the selected-versus-null rank into an empirical p-value, then calibrate the final maximum over experts with the existing conformal cascade.

This is stronger than scoring every candidate on the same pixels: selection noise cannot confirm itself, and textured images define their own spectral background.

Gate: p-values must be approximately super-uniform on every locked negative stratum. A positive-only score improvement without null calibration does not advance.

H4: content-whitened multichannel matched filter

The current template correlations normalize global energy but do not model the image-specific complex noise covariance around carrier bins. Estimate local power spectral density and the full RGB or LMS cross-channel covariance from neighboring noncarrier frequencies, whiten the residual, and then apply the frozen complex carrier template. Learn at most one generalized-eigenvector channel projection per expert on training data; Green, Lab, and OKLab are baselines rather than separately selected hypotheses.

Gate: whitening must improve within-parent carrier margin in high-texture positives and reduce the heavy null tail on natural resampling aliases. It must not depend on scene category or expose RGB thumbnails to a learned model.

H5: repeated unknown-payload consistency

Cross-image phase lock can fail when every image carries a different payload or key. The watermark may still repeat one internally consistent phase or sign codeword across distant patches of the same image. Estimate a short complex phase/sign vector independently in each nonoverlapping patch, allow one global cyclic phase and polarity nuisance parameter, and test patch-to-patch agreement. Do not require the codeword to match another image.

This generalizes the existing quadrant minimum: it tests a structured vector and its parity or low-rank consistency rather than one scalar correlation. Use synthetic random codewords and phase-scrambled equal-power fields to mutation- test that the statistic recognizes repetition, not merely periodic energy.

Gate: agreement must transfer to new source images and remain absent from periodic graphics, demosaicing grids, JPEG blocks, and renderer lattices. If no same-image consistency exists in low-texture oracle positives, reject the payload-repetition model.

H6: calibrated encoder-state mixtures

One provider may deploy several encoder keys, payload families, strengths, or model epochs. Fit a small mixture only on matching-oracle positives or causal pairs, freeze every component before evaluation, and expose each component as a separately identified expert. The union uses search-corrected conformal p-values, not per-component thresholds chosen from the test set.

The earlier OpenAI multi-direction codebook does not justify this branch: it was trained on assertion-labeled exports and failed fresh specificity. Reopen the hypothesis only with the P1 OpenAI prerequisites. For Gemini, test whether the known first and orthogonal carrier directions remain separate on official pixel positives.

Gate: a mixture must rescue locked positives without increasing the familywise false-positive bound, and every retained component must recur in a later time window. Singletons and post-test clusters are discarded.

H7: residual-only equivariant network

If H1-H5 expose a causal statistic but a hand-built aggregation loses recall, train a compact network on high-pass residual fields with explicit translation, rotation, scale, and codec transformations. Give it complex carrier maps or steerable-wavelet responses, not RGB content. Enforce source-matched pairs, leave one encoder state and one transform family out, and keep the classical expert as an independent confirmer.

Gate: the network must improve the paired locked test and the unseen-transform test at the same low-FPR boundary. Saliency must remain on residual carrier structure; a content or border shortcut kills the branch.

H8: optional multi-image evidence

A user may possess a set of outputs from one generation session. A separate batch detector can aggregate weak phase or codeword evidence across diverse images after content-hash grouping. It may detect a shared encoder state that is too weak in one image, but it must return batch-level confidence and must not relabel every member as individually detected.

Gate: define a minimum number of content-diverse parent groups, evaluate whole sessions as the independent unit, and challenge same-renderer negative batches. If gains disappear after session and content balancing, the result is provider attribution rather than watermark detection.

Hypothesis execution policy

Run H1, H3, and H4 first for Gemini because they directly target the measured resize, texture, and large-image failure modes. Run H2 and H5 as the next mechanism probes because they can detect payload-varying structure without a shared cross-image template. H6 and H7 require the new oracle-positive corpus; for OpenAI they additionally require the authorization or causal-pair boundary in P3. H8 is a separate product mode, not a substitute for single-image qualification.

At most one hypothesis family may consume a locked partition. A failed locked test closes that partition for model selection and opens a new version only after a fresh holdout exists. This is the cost of being creative without turning repeated hypothesis search into an invisible false-positive generator.

2026-08-14: split-confirm lattice synchronization result

The first H1/H3 prototype is implemented in synthid_affine_lattice_probe.py. It samples strong complex harmonics from 16 nonoverlapping checkerboard patches, selects period candidates on eight patches, and confirms the chosen lattice on the other eight after correcting every patch for its global origin. Five separated zero-rotation period candidates then enter an amplitude-aware reranker. Cyclic phase and period are selected from one patch group and scored on the other. The final full-image template correlation preserves the observed phase; a separately reported cyclically registered score is diagnostic only.

The separation matters. A broad coherence search at 0.25-pixel spacing selected harmonic aliases such as 10.5 instead of the true 12.8-pixel period. Refining the grid to 0.1 pixels recovered the exact expected periods in all 20 transformed positive views: 12.8, 14.4, 17.6, and 19.2 pixels for scale factors 0.8, 0.9, 1.1, and 1.2. A candidate-count mutation from five to one selected the same periods and produced the same template scores in all 20 views. The grid resolution, not the amplitude reranker, caused this recovery; amplitude remains an independent confirmation statistic. Cyclic registration was also rejected for the ordinary scale route. It increased scores on resize-interpolation lattices, whereas the phase-preserving correlation kept those aliases negative.

One provisional conjunction used the already frozen fixed-template threshold 0.17357069773071196, split-confirm lattice coherence of at least 0.52, and split-confirm amplitude of at least 0.20. These latter two constants are research thresholds, not runtime calibration. Results were:

Challenge	Positive result	Negative or cross-provider result
Native matching-oracle pilot	5/5	0/130 matching-oracle Picsum controls
Four-scale development matrix	20/20 transformed views from 5 parents	0/200 views from 50 oracle-negative parents
Four-scale locked matrix	20/20 transformed views from the same 5 positive parents	0/800 views from 200 previously unused oracle-negative parents
Native large official positives	7/7	Not a new negative challenge; the frozen large-v1 controls remain the applicable evidence
Source-disjoint provider-labeled test	295/455 Google-labeled candidates	0/469 OpenAI-labeled candidates
Broad non-Google challenge	Not applicable	3/276 scored rows, plus 6 unsupported small images

The four transformed views of one parent are correlated observations, so the scale rows establish transform consistency for 5 positive and 250 negative parents, not independent sample sizes of 20 and 1,000. Across the locked scale matrix, the phase-preserving template score had positive minimum 0.3603 and negative maximum 0.1716; split coherence had positive minimum 0.5344 in the development matrix and locked negative maximum 0.5028. All 12 independent matching-oracle native positives, five ordinary-size and seven large, selected period 16 and crossed the provisional conjunction.

The three broad-challenge crossings are not established false positives. Two carry TC260 AIGC provenance and one carries Samsung Galaxy C2PA provenance; all three independently crossed the earlier fixed carrier detector. They show that the measured carrier family is not safe for provider attribution. They may be another compatible watermark deployment, Google-backed editing inside a different product, or a shared renderer lattice. Only a matching pixel oracle or source-matched counterfactual can distinguish those explanations.

The cross-provider result is nevertheless useful: the Google codebook plus the new synchronization path accepted none of 469 OpenAI-labeled test images. This supports a different OpenAI payload or carrier family rather than a uniformly weaker copy of the measured Google code. It does not provide a local OpenAI detector; the official OpenAI pixel verifier remains the only production route.

Two related hypotheses were rejected. Phase-only cyclic codeword agreement retained hard non-Google carrier rows, so amplitude cannot be discarded. A period-8 decoder made by antialiasing the 16-by-16 codebook into 8-by-8 templates with area, linear, cubic, or Lanczos kernels also failed: its best Google-versus- control AUC was 0.567, and split amplitude remained near chance. Period 8 is therefore still an unsupported state, not a threshold adjustment.

This result promotes split-confirm synchronization to the next Gemini research candidate, but not to runtime. The positive oracle set remains small, the provider-labeled corpus is not a clean oracle, and the amplitude reranker currently supports isotropic scale at zero rotation only. Rotation, crop, translation, perspective, JPEG/WebP, independent temporal positives, and a fresh clean negative source must pass before a production expert is calibrated.

2026-08-14: content-whitened matched-filter result

H4 was implemented as a complex multichannel matched filter on the 16 strongest template harmonics. For each harmonic and patch, a ring of neighboring noncarrier bins estimates the image-specific complex covariance of Green, Red-minus-Green, and Blue-minus-Yellow. A fixed ridge regularizes that covariance before the template and observed carrier vectors are whitened. Period is chosen from the five H1 candidates on one checkerboard of patches; the reported score is the smaller whitened cosine on the selection and disjoint confirmation checkerboards. This retains the P1 separation and charges the candidate search to the observed negative tail.

The implementation passed synthetic carrier, independent-noise, resize, and cyclic-phase mutations. It also found a real synchronization failure that the amplitude reranker missed: under a broad native search, one of five official ordinary-size positives selected a period-24 harmonic alias. Choosing among the same frozen candidate set by whitened selection score changed that result to period 16, and all five ordinary-size positives then selected period 16.

The matched-filter score itself did not pass its preregistered improvement gate:

Challenge	Whitened positives	Whitened negatives	Whitened margin	Existing template margin
Native pilot	5/5, range 0.7006-0.8243	130 controls, maximum 0.4797	0.2209	0.2650
Locked 0.8 resize	5/5, range 0.4268-0.5994	200 fresh-parent controls, maximum 0.2165	0.2103	0.2420
Native large positives	7/7, range 0.5401-0.6017	No new negative partition consumed	Not estimated	Not estimated

The different score scales make the absolute margins diagnostic rather than a shared operating threshold, but both locked comparisons point in the same direction: whitening separated the measured cohorts yet did not improve the boundary over the phase-preserving spatial template. H4 is therefore rejected as a replacement detection statistic. Its candidate-selection role remains a useful H1 synchronization refinement and must be calibrated as part of that search, not added as another post-hoc threshold.

The full complex code also sharpened the interpretation of the non-Google challenge. The three earlier TC260- or Samsung-provenance crossings scored 0.4764, 0.6666, and 0.7746; two match the whitened code more strongly than the weakest official large positive. They cannot be dismissed as generic period-16 image structure. They remain compatible watermark-family detections, Google-backed edits, or a shared implementation until a pixel oracle or causal counterfactual labels them. Provider attribution from this code is still invalid.

Finally, the ten OpenAI-labeled rows with the largest earlier Google-template scores reached at most 0.2234 under the whitened Google code. This selected challenge is not a sensitivity estimate and does not create an OpenAI detector, but it is inconsistent with the hypothesis that those hardest OpenAI rows carry the same code at merely lower amplitude. The next local OpenAI experiment still requires authorized oracle labels or watermark-on/off pairs from which to learn its own carrier and codebook.

2026-08-14: repeated unknown-codeword result

H5 was tested without requiring the observed phase vector to match the known Google template. The codeword is estimated from one checkerboard of patches; the score is the magnitude of its complex weighted inner product with the independently estimated codeword from the other checkerboard. The 16 harmonic coordinates remain preregistered from the carrier template, but the phase and sign values are learned within each image. A global phase or polarity cancels in the cross-product.

The initially proposed free cyclic-shift nuisance was rejected. Selecting a shift on alternating harmonics and confirming it on the other harmonics gave official-positive scores of 0.3651-0.6739, but the 130-control maximum was 0.4698. A fixed zero-relative-shift variant also overlapped. Natural periodic phase supplied enough shift aliases that search-corrected confirmation did not separate the cohorts.

After global patch-origin correction, no relative shift is expected. The final preregistered variant therefore uses all 16 harmonics, performs no shift or harmonic selection, and confirms only across the two patch groups. It produced:

Challenge	Official or oracle-positive range	Control maximum	Margin
Native fixed period 16	5/5, 0.4498-0.7291	130 controls, 0.2879	0.1619
Locked 0.8 resize at period 12.8	5/5, 0.3922-0.6713	200 fresh-parent controls, 0.2381	0.1541
Native large fixed period 16	7/7, 0.4044-0.6485	No new negative partition consumed	Not estimated

Synthetic mutation tests confirmed the intended invariance: the score accepted the repeated random carrier, rejected independent equal-power noise, survived isotropic resize after period canonicalization, and stayed high after a global cyclic image translation even though the known phase-preserving template score collapsed.

The three TC260- or Samsung-provenance compatible carriers scored 0.3994, 0.4541, and 0.6052. Their within-image codeword repeat is therefore also inconsistent with an accidental single global Fourier spike. This strengthens the compatible-signal-family interpretation without identifying a provider. The ten OpenAI-labeled rows selected for their earlier Google-template scores reached at most 0.2568 at fixed period 16, below the native-control maximum. That selected result is not a full OpenAI screen, but supplies no evidence that the same period-16 harmonic support carries a different repeating OpenAI payload.

H5 advances as an independent research confirmation, not a runtime threshold. The current evaluation supplies the correct period rather than charging the full period search to this statistic, has only 12 independent official positive parents, and reuses the available Picsum negative source. It must next survive the broad H1 candidate search, the remaining scale and codec matrix, a second clean negative family, and independent temporal positives. A free-shift version is not reopened for the ordinary scale route; translation and crop belong in a separately calibrated expert.

2026-08-14: complex cyclostationary result

H2 was implemented as the normalized complex spectral-correlation matrix F(k + alpha) * conjugate(F(k)) across all nine Green/opponent-color channel pairs. The eight strongest preregistered template harmonics supplied alpha. For every carrier shift, six one-bin neighboring shifts formed a same-radius same-image null. Complex matrices were averaged separately over the selection and confirmation checkerboards before taking their Frobenius norms, so content phase had to repeat across distant patches to count.

Synthetic mutations verified that this was a genuine cyclostationary test. An amplitude-modulated equal-power random field produced positive carrier-minus- neighbor contrast on both checkerboards; independent noise and an additive periodic carrier did not. The latter mutation distinguishes H2 from the H1/H5 additive-codeword family.

The native oracle pilot failed decisively:

Statistic	Five official positives	130 controls	Positive minimum minus control maximum
Selection contrast	-0.0027-0.0342	maximum 0.0593	-0.0620
Confirmation contrast	0.0067-0.0337	maximum 0.0608	-0.0542
Joint contrast	-0.0027-0.0337	maximum 0.0593	-0.0620

H2 is therefore rejected without consuming another locked scale partition or searching frequency bands, harmonic subsets, color spaces, or null offsets on the observed controls. The measured Google signal supports a repeated additive or content-adaptive codeword, not a detectable global multiplicative cyclostationary modulation. This result does not constrain an unknown OpenAI carrier, but the OpenAI branch still lacks the oracle evidence required to fit one without provider leakage.

2026-08-14: fixed-period codec and crop pilots

The H1/H4/H5 probe gained explicit in-memory JPEG and top-left crop transforms so the same decoded parent pixels enter every paired comparison. These pilots use the five ordinary-size official positive parents and split the 130 native controls into 50 development and 80 locked controls. The transformed views are not new independent positive evidence.

Lossy JPEG separates the useful statistics. Whitened and unknown-codeword scores no longer maintain a positive development margin even at quality 95, so they are unsupported for the lossy-codec route. The phase-preserving spatial template remains usable at the known native period. Thresholds were frozen after the 50-control development split and then applied unchanged:

Transform	Frozen template threshold	Repeated positive views	Locked controls	Locked score margin
JPEG 95	0.12	5/5	0/80	0.1054
JPEG 90	0.09	5/5	0/80	0.0727
JPEG 80	0.065	5/5	0/80	0.0257

The declining margin is monotonic and leaves JPEG 80 especially fragile. These constants are research thresholds, not a runtime calibration: they use only five positive parents, one negative family, fixed period 16, and one OpenCV JPEG implementation. JPEG 100, chroma-subsampling variants, WebP, a second encoder, codec-after-resize combinations, and independent positives remain unqualified.

Cropping shows the opposite phase behavior. Removing 5% or 10% from the top and left collapses the unregistered template score, while cyclic phase registration recovers the carrier. A crop-only conjunction was frozen at registered template at least 0.20, split coherence at least 0.55, and split amplitude at least 0.20:

Transform	Repeated positive views	Locked controls	Registered-template locked margin
5% top-left crop	5/5	0/80	0.2275
10% top-left crop	5/5	0/80	0.2378

The unknown-codeword H5 statistic independently retained positive margins of 0.0872 and 0.0839 in those two cells. This supports a separate native period-16 crop/translation expert. It does not reopen cyclic registration in the ordinary scale route, where the same nuisance parameter was already shown to amplify interpolation aliases. Arbitrary crop sides and offsets, one-pixel translations, combined crop-plus-codec edits, rotation, screenshots, broad negative sources, and new positive parents remain before qualification.

2026-08-14: fresh-source failure and frozen v3 confirmation candidate

The unchanged runtime experts were challenged on the first 3,000 decoded- pixel-unique files from the official Open Images test split. The image-id list was fixed before download and has SHA-256 9b83bc221dd70823f39a9ad62b16c1dcc1f9c449f17c9ea3f6c9a51b581e0c1a; the ordered pixel-file hash list has SHA-256 2c49ccdec79ec662f202670b9e4c5ee0dfdd80d5747929dd09795d945b3a22a6. This source-fresh challenge rejected both existing precision claims. The fixed expert accepted 5 of its 211 supported images at its shipped 0.1735706977 threshold. The registered expert accepted 4 of 2,997 supported images at its shipped 1.0 threshold. These are false-positive measurements, not clean-image labels inferred from detector misses.

The fixed v3 research candidate is frozen at 0.28. It is above the largest fresh crossing, 0.2592123, and below the weakest of 12 source-diverse native Google positives, 0.2978951. That separation selected the threshold, so the 3,000-image set is development evidence for v3 and cannot validate it. The candidate must remain unchanged on a second untouched source split.

The registered v3 research candidate preserves the unchanged registered base decision and adds an independently split phase confirmation. Selection and confirmation use disjoint checkerboards of patches. The confirmation requires period at least 10.0, joint harmonic coherence at least 0.30, and nonnegative joint cyclic amplitude. Three narrow period ranges add the following measured alias controls:

Registered period	Additional frozen gate
18.0 <= p < 18.6	held-out unknown-codeword confirmation at least 0.13
18.6 <= p < 20.0	joint harmonic coherence at least 0.40
20.0 <= p < 22.0	held-out unknown-codeword confirmation at least 0.02

The rule rejected all four fresh registered crossings while retaining 147 of 148 base-accepted positives from 226 source-diverse transformed parents. A separate dense transform matrix applied 35 scales from 0.65 through 1.50 in 0.025 increments to 12 native positive parents; it retained all 359 views that the unchanged registered base accepted. At the hard-negative period 18.28, all 11 base-accepted positive views passed, while the natural-image crossing failed the held-out-codeword gate. Derivatives are not counted as independent positive evidence. The exact candidate constants live in scripts/synthid_affine_lattice_probe.py and are frozen before the second negative holdout is acquired.

An explicit period-8 rescue is still rejected. A six-gate candidate selected on earlier controls accepted 7 of these 3,000 fresh controls, so scale 0.5 has no qualified local route.

2026-08-14: registered-v3 holdout qualification and weak-signal rescue

The registered confirmation rule above was frozen unchanged before a second 3,000-file Open Images test-split cohort was decoded. Its image-id list has SHA-256 91e1f19f559a432f61c5eb1f16797d40d706170c43094bc7d4cb9506309b662a; the ordered pixel-file hash list has SHA-256 651f959689e7043989ded09fa4aece5cea2592e7b69e371cc97c35f581d9f57d. All 3,000 decoded hashes were unique and none overlapped the first cohort. The unchanged registered-v2 base produced no threshold crossings among 2,996 supported images, so registered-v3 also produced 0/2,996. Together, the two Open Images cohorts give registered-v3 0/5,993 false positives after the confirmation rule was applied, with a one-sided 95% zero-error upper bound of about 0.050%. They are source-disjoint samples from one source family, not two independent acquisition mechanisms.

A second-family challenge used the first 3,000 sorted COCO val2017 files. The newline-delimited file-name list has SHA-256 7a2e5f47f6c775f0de839b8270ee0874526a8d0984c66cd7940e1a04fc89f4f5; the ordered per-file SHA-256 manifest has SHA-256 b1b0c73891b6d5b49a7817c9bd0a47309a424e437ba36fabed7bd9077dfeda54. All 3,000 pixel hashes were unique. The unchanged registered base produced 0/2,366 crossings, with maximum score 0.932964706; registered-v3 therefore also produced zero. COCO predates this confirmation experiment and is not a fresh acquisition, but it is a distinct natural-image source family. Across both Open Images cohorts and this COCO challenge, registered-v3 has 0/8,359 supported-control crossings. This qualifies the precision-first positive route; it does not turn a miss into proof of absence.

The fixed 0.28 candidate failed the second holdout and is rejected. It accepted one of 213 supported images, at score 0.322542963. The runtime keeps fixed-v2 only as an explicitly requested diagnostic and never unions it with a production positive route.

Split confirmation was also tested as a recall rescue below the registered base threshold. The first relaxed rule recovered 16 of 76 source-diverse base misses but accepted 94 of the 312 most dangerous first-cohort controls, so repeated phase alone is decisively insufficient. A stricter H4/H5 conjunction required matching spatial and spectral periods, base amplitude margin at least 0.4, high-band margin at least 0.3, the frozen confirmation gates, and a minimum of content-whitened template match, all-harmonic repeated-codeword match, and split cyclic amplitude at least 0.225. The threshold is above the first- cohort maximum 0.222382188. It added only 6 of 224 scored source-diverse positives and 8 of 432 dense-scale views, while accepting none of the second- cohort candidates and none of the scored COCO candidates. That small recall gain does not justify another runtime branch: it was selected on exposed development positives, adds a substantially more expensive whitening pass, and lacks new positive parents. It remains a recorded research candidate rather than being hidden inside registered-v3.

The runtime implementation was then checked on the transformed pixels rather than the original source paths. Of 226 source-diverse transform assignments, 224 remained large enough for the split-patch research scorer. Registered-v3 detected 147/224 overall: 0/40 at 0.5x and 147/184 (79.9%) from 0.65x through 1.5x. The v2 base accepted 148 of the latter 184 and v3 confirmation rejected only one base crossing. Failure attribution on the 184 supported non-half-scale views was 10 period-winner mismatches, 17 amplitude failures, 9 high-band failures, one confirmation failure, and 147 detections. This production-parity measurement replaces the earlier accidental source-path check; transform labels attached to untransformed files are not sensitivity evidence.

H3 then supplied a genuinely content-matched null. For each registered period, the 16 fixed neighboring periods at offsets from plus or minus 0.35 through 2.0 pixels are scored on the same two checkerboards. The candidate statistic is the smaller of the selection- and confirmation-group target coherence minus that group's largest off-period coherence. Combining the two Open Images development cohorts froze a recovery rule at registered base score at least 0.8, the existing v3 confirmation gates, and same-image joint excess at least 0.05695968150082637. It rescued 7 of 36 sub-threshold transformed views and accepted 0/262 prescreened candidates from the first 3,000 COCO controls. The untouched final 2,000 COCO files contained 1,598 registered-supported images and seven near-threshold candidates. Four candidates were too small to form two independent 256-pixel patch groups; all three scored candidates had joint excess at most zero. The frozen rule and registered-v3 both accepted 0/1,598. This passes the new negative-side test without changing the threshold. The candidate remains outside runtime because all seven recovered positives influenced feature selection and no new positive parent has yet confirmed the gain.

The same null did not justify relaxing the registered requirement that the spatial reconstruction winner equal the strongest spectral-period candidate. A full rescore of the two 3,000-file Open Images cohorts retained 469 plausible mismatch candidates after amplitude and high-band prefilters; 212 also passed the frozen split confirmation. Their same-image joint-excess maximum was 0.1762448251. Only five of the ten non-half-scale positive mismatch records passed the same mechanism prefilters, and their maximum was lower at 0.1464860382. No zero-control same-image threshold retained a positive. A post-hoc product of amplitude margin, high-band margin, and coherence could retain two positives above the observed control maximum, but it was selected from many combinations and offers too little gain to open another holdout. Spatial/spectral winner equality therefore remains mandatory.

2026-08-14: content-adaptive patch-phase recovery candidate

The official SynthID-Image paper describes a content-dependent learned watermark pattern, so a single whole-image template cannot be assumed to be the complete decoder. A new research statistic therefore treats the frozen tile as a synchronization basis rather than as a fixed global payload. At the registered period, each 256-pixel patch is folded independently, every cyclic tile shift is standardized against that patch's own shift distribution, and the strongest half of the selection checkerboard chooses one phase. The score is the smaller of that phase's trimmed standardized response on the selection and disjoint confirmation checkerboards. Filenames, metadata, and scene labels do not enter the statistic.

The research-only recovery rule was frozen at matching spatial and spectral periods, the existing split confirmation, both registered gate margins at least 0.45, one gate margin at least 1.0, and joint patch-shift score at least 2.5. It recovered 5 of 18 scored amplitude/high-band failures from the 224-view transformed-positive audit. Added to registered-v3 alone, that would move the measured non-half-scale result from 147/184 to 152/184 (82.6%). It accepted none of 88 gate-failure candidates in the first 3,000-image Open Images development cohort and none of 334 previously scored control candidates; those sets overlap and are not added as independent denominators. The rule and threshold were recorded before reading a 350-image Picsum feature-holdout. All 350 decoded rasters were unique. The frozen rule accepted none; only two controls reached its mechanism pre-gates, and their maximum joint patch-shift score was 0.7029283 against the 2.5 threshold. The zero-error one-sided 95% upper bound is about 0.85% for this one-source feature-holdout. The candidate remains outside runtime until new positive parents pass unchanged.

That positive test then failed on a genuinely different carrier cohort. Seven large Gemini images with independent official pixel-positive verdicts, none of whose source paths appeared in the 224-view transform audit, were resized by 0.65, 0.70, and 0.75. Registered-v3 accepted 0/21 and the frozen patch-phase recovery added 0/21. The base amplitude was often strong, but the RGB high-band margin became negative on most views and several period searches selected an octave alias. The branch is therefore a real recall improvement for the measured registered carrier, but not a universal cross-epoch solution. A scale-registered form of the independently validated large-v1 opponent-color signal is the next hypothesis; weakening the existing high-band gate is not.

That hypothesis recovered the entire failed cohort. A signed spectral curve in the Red-minus-Green and Blue-minus-Yellow planes searched periods 10.0 through 12.5 in 0.05-pixel steps, retained three separated candidates, canonicalized each to the frozen period-16 tile, and reranked by the minimum of fixed RGB, Red-minus-Green spatial, and Blue-minus-Yellow spatial agreement. Thresholds were frozen at 0.16, 0.60, and 0.55, respectively. Automatic period selection recovered the exact injected period on all 21 views and accepted 21/21, versus registered-v3's 0/21. It accepted 0/21 symmetrically resized views from seven large Open Images controls; the maximum control decision margin was 0.9462124. These positives and controls selected the gates, so this is development evidence. The candidate was frozen before a 350-image Picsum feature-holdout. It accepted 0/350 decoded-pixel-unique controls there; the maximum decision margin was 0.7778322. This negative-side result passed unchanged.

The broader challenge rejected that unrestricted form. Extending the search to periods 7.5-14.5 recovered 63/63 lossless views from the same seven official parents at nine scales from 0.5 through 0.9, but accepted 605/2,997 and 587/2,996 controls in two Open Images cohorts. The crossings concentrated at periods 8 and 12.8 and tracked JPEG sampling. A 3,964-image COCO challenge added one period-10 crossing below one megapixel, while 3,772 heterogeneous non-Google/non-OpenAI Spaces controls added 28 crossings, again only at periods 8 and 12.8. The unrestricted branch is a codec/resize-lattice detector and is rejected.

A symmetric codec test separated lossless scale recovery from recompression. At scales 0.5, 0.625, and 0.75 the raw views passed 21/21, while JPEG-95, JPEG-85, and WebP-95 passed 0/63. At the supplied target period, half-scale JPEG views inverted the opponent-color polarity and the other lossy views largely lost spatial agreement. This is not a period-search-only failure.

The production candidate was then frozen at 1-10 megapixels, minimum side 768, selected period 8.1-12.0, and the original 0.16, 0.60, and 0.55 gates. An initially added spectral/spatial period-agreement condition was removed when the real API path showed that a harmless 11.0 spectral alias reduced 0.55x recall from 7/7 to 1/7; its only reserve-control counterpart was below the already frozen geometry floor. The revised rule was fixed before a separate 1,000-image Picsum holdout. It accepted 0/1,000, with maximum unrestricted margin 0.7953734, and the production detect_synthid seam detected 42/42 lossless views across scales 0.55-0.75. This bounded expert now runs after registered-v3 abstains as synthid-periodic-tile-opponent-registered-v1.

2026-08-15: period-8 codec veto and lossless 0.5x recovery

The original period exclusion conflated two observations: the 0.5x watermark and the native JPEG block lattice both select period 8. Direct template polarity did not resolve the confound. JPEG-95 and JPEG-85 half-scale positives produced a strong inverted opponent template, but random JPEG 4:2:0 controls produced the same profile. A matched resize-plus-JPEG residual experiment used 175 control parents for development, 175 disjoint parents for holdout, three JPEG qualities per parent, and leave-one-positive-parent-out checks. The best ridge variant accepted 19/21 fitted positive views but also 86/525 holdout controls; the variants with fewer control errors lost most or all positive parents.

A nonlinear block-DCT follow-up evaluated 80 target-derived statistics in RGB, Green, opponent-color, and YCbCr views on the same matched transform. Its best zero-holdout-error statistic retained only 2/21 JPEG positives. The highest-AUC zero-error opponent statistic reached about 0.735 but retained 0/21 at the development maximum. Neither JPEG recovery hypothesis enters runtime.

The successful discriminator is instead a codec veto around the unchanged opponent carrier gates. For each decoded opponent plane, it measures the mean absolute edge energy on native 8-pixel boundaries relative to the median of non-block phases. All seven lossless 0.5x positives had both Red-Green and Blue-Yellow ratios no greater than the frozen 1.05 ceiling. The ceiling rejected 600/600 development period-8 codec crossings, then 582/582 crossings in the nonoverlapping Open Images holdout, 601/601 in the feature-unseen frozen reserve, and 7/7 in the heterogeneous Spaces challenge. An additional 350 controls transformed through the identical lossless 0.5x resize produced zero base crossings before the veto.

The production fallback now accepts periods 7.9-12.0 and requires the veto through period 8.1. The public API detects 7/7 scale-0.5 views and 49/49 across the measured 0.5-0.75 lossless matrix from seven official pixel-positive parents. This is a bounded lossless extension, not evidence of JPEG robustness or arbitrary scale coverage.

The provider challenge is evidence for a shared family, not attribution. The unrestricted statistic accepted 95/7,254 OpenAI files with an explicit SynthID assertion, all at periods 10.70-10.80, and 4/287 same-lineage files without an assertion. The latter label is not an oracle-negative watermark verdict. The 95 decoded rasters were unique and had zero overlap with the 94 period-8 candidate hits, producing a 189/7,259 union. However, JPEG-95 and WebP-95 each reduced the new component from 95/95 to 0/95. It remains compatible with either a complementary carrier epoch or a geometry-linked export lattice. The runtime result remains provider-neutral, and the official OpenAI verifier is still the only qualified provider-wide OpenAI route.

P3: productionize OpenAI without pretending the local route works

The near-term production route is the official pixel verifier. Remaining live and deployment qualification:

• complete the approved SDK transport set beyond the finished positive PNG smoke, covering JPEG, WebP, and pre-established positive and negative expectations with pixel-identical, metadata-free files;
• malformed file, 50 MiB boundary, timeout, disconnect, 400, 404, 429, 5xx, unexpected response, duplicate SynthID result, and missing SynthID result qualification against recorded or contract-safe fixtures;
• deployment-level bounded backoff and a circuit breaker for transient API failures, without representing transport failure as a watermark verdict;
• confirmation of data-retention, access-control, audit-log, and user-consent requirements for the deployment environment;
• secret-safe logs that retain request ids, byte counts, timings, and status but never image bytes, credentials, user paths, or decoded-pixel hashes.

The local transport now fixes both the default SDK and per-request timeout at 120 seconds, sets automatic retries to zero so one acknowledgement cannot retransmit media, and omits the decoded-pixel hash from request logs. Contract tests cover exact and over-limit file sizes, caller cancellation, malformed responses, all documented HTTP classes, timeout, and connection failure. A structured error preserves status, API error code, request id, Retry-After, and whether an explicit retry is appropriate. One credentialed positive PNG smoke has now passed. Remaining format qualification, deployment retention approval, and an operational circuit breaker remain environment-level work; the research harness itself does not perform live uploads.

A local OpenAI detector reopens only after one of two prerequisites exists: written authorization to obtain non-adaptive pixel-verifier labels at research scale, or source-matched watermark-on/watermark-off pairs. The first local epoch then starts from those causal residuals and uses same-provider negatives. No additional model is selected on C2PA assertions, filenames, OpenAI export geometry, or the already exhausted provenance-labeled cohorts.

Updated local-detector execution order

The local program now advances through evidence gaps rather than adding more unbounded feature searches.

1. Freeze the current routed Gemini bank, calibration artifacts, parent-group manifests, and support reasons as the reproducible baseline. Finish the active fine/anisotropic registration experiment against a fresh negative holdout and either promote it as a separately versioned expert or reject it before opening another synchronization branch.
2. Run the production router, not isolated expert functions, over the complete available size, scale, aspect, orientation, and codec matrix. Group results by parent and cluster misses and abstentions by measured carrier state. This produces the actual coverage backlog.
3. Address Gemini clusters one at a time. Use full reciprocal-lattice registration first, select-confirm same-image null and repeated unknown codeword confirmation second, and codec-specific residual experts third. Every branch selects on one region group, confirms on disjoint regions, and receives its own multiple-search null calibration.
4. Do not train a local OpenAI classifier from renderer style or provenance assertions. Existing verifier-confirmed cases may validate whether a candidate reaches the known pixel signal, but the training corpus requires source-matched watermark-on/off pairs or written authorization for a non-adaptive research label set. Same-provider hard negatives are mandatory.
5. Once that OpenAI prerequisite is met, learn causal residual differences, cluster stable carrier states across model and time windows, and expose each recurring state as a separate versioned expert. Test reciprocal-lattice, multichannel complex matched-filter, and within-image unknown-codeword families before any residual-only learned model.
6. Qualify both provider banks through one locked and later temporal matrix. A positive names only the provider scope proved by exclusivity data; a miss is not_detected only inside sensitivity-certified strata and otherwise remains indeterminate or unsupported.

P4: run one symmetric robustness matrix

Every promoted expert and both official-verifier integrations receive the same parent-grouped challenge:

Axis	Frozen views
Codec	lossless PNG, source JPEG, JPEG 100/95/90/80, WebP lossless/95/80, 4:4:4 and 4:2:0 where controllable
Scale	0.5, 0.625, 0.75, 0.8, 0.9, 1.0, 1.1, 1.2, 1.333, 1.5, and 2.0
Geometry	supported size-bin boundaries, modulo-period edges, portrait, landscape, square, and extreme aspect ratios
Spatial edit	1-10% crop, one-pixel and multi-pixel translations, screenshot round trip, and 90/180/270-degree rotation
Pixel edit	mild blur, sharpen, denoise, gamma, contrast, saturation, and color-profile conversion
Content	photo, face, text-heavy, illustration, flat graphic, low texture, high texture, and alpha

Transforms are applied identically to positives and controls. Scores and verdicts are aggregated by parent, not by treating eleven derivatives as eleven independent images. Unsupported cells are recorded as such; they do not lower the denominator silently. Robustness work stops when the signal is no longer identifiable at the required false-positive boundary.

P5: use release-grade statistical gates

Freeze the following gates before the new corpus is scored:

• the one-sided 95% upper confidence bound for false-positive rate must be at most 0.1% overall and at most 0.5% in every critical negative stratum;
• the one-sided 95% lower confidence bound for sensitivity must be at least 90% in every scope advertised as detection-supported;
• a local not_detected absence claim needs a much stronger 99% sensitivity lower bound in every advertised stratum; otherwise a miss remains indeterminate;
• temporal, model-family, surface, geometry, and codec results are reported separately, with no claim wider than the weakest passing stratum;
• every selected threshold must remain unchanged on a newly acquired temporal holdout and on a second negative source family;
• provider attribution requires cross-provider exclusivity data. Until that passes, the local result names the signal family but not its generator.

Confidence intervals, operating thresholds, model hashes, calibration hashes, split hashes, and code revision go into one machine-readable qualification report. The report fails closed on missing records, duplicate parent groups, NaN or infinite scores, unknown detector ids, or a changed artifact hash.

P6: harden runtime and operations

Before release:

• benchmark wall time and peak memory at every supported size boundary; keep large-image folding and template subtraction stripe-bounded, and freeze numeric p95 latency, peak-memory, and concurrency budgets for the target deployment before qualification;
• fuzz corrupt, truncated, decompression-bomb, unusual color-mode, alpha, and oversized images; errors must be bounded and must not become negatives;
• prove deterministic scores across supported platforms and dependency pins, or version platform-specific calibration explicitly;
• run shadow mode before user-visible verdicts, collecting only consented, privacy-safe aggregate score histograms, support rates, latency, errors, and detector-version counts;
• alert on score-distribution drift, support-rate changes, remote outcome drift, error-rate spikes, and latency regressions;
• keep the previous detector and calibration artifacts available for immediate rollback, and make every result identify which version produced it;
• require the applicable dependency lock, security scan, Ruff, formatting, Pyright, unit, integration, mutation, packaging, and full corpus gates to be green. An untriaged production dependency advisory blocks release.

Execution order and stop/go checkpoints

Milestone	Deliverable	Go condition
M1	Unified result contract and expert registry	No miss, error, unsupported input, or remote failure can be presented as clean
M2	Immutable Gemini and OpenAI qualification manifests	Parent-group split audit and all ingestion mutations pass
M3	Gemini fixed and large requalification	P5 precision and sensitivity gates pass on locked and temporal data
M4	OpenAI SDK/API qualification	Approved live transport matrix passes, privacy boundary is accepted, operational failures abstain
M5	Optional registered, codec, and crop experts	Each expert independently passes P5; failed cells stay unsupported
M6	Shadow deployment	No unexplained drift or SLO regression through the prespecified observation window
M7	Production release candidate	All code, corpus, security, documentation, rollback, and versioning gates pass

The immediate work queue is M1, freezing and auditing the current routed baseline, completing the active Gemini fine/anisotropic candidate decision, finishing the 3,000-image fresh Gemini large negative acquisition, and running the router-level all-resolution matrix. In parallel at the planning and data provenance level, identify a permitted source of OpenAI watermark-on/off pairs or obtain written authorization for a non-adaptive research label set. The first credentialed OpenAI positive PNG smoke is complete; further live API calls are transport qualification, not local-detector training. New feature search starts only from a measured coverage cluster and consumes a fresh holdout once.

2026-08-15: OKLab random-codeword replacement probe

A pixel-only probe tested the hypothesis that the measured periodic residual is only the carrier for a repeated coded message. It compared independent high-pass OKLab noise, periodic-comb suppression, random periodic injection, and complete replacement of the estimated period-16 residual with a magnitude-matched random-phase tile. The four inputs were the cleared public Google-oracle-positive originals already in the tracked research corpus.

Independent noise was inefficient. At the stronger tested allocation, with three times more RMS assigned to each chroma channel than to luminance, the median fixed-carrier score changed by only -0.0219 while median PSNR fell to 32.56 dB. Comb suppression alone changed the median score by -0.1962 at 58.05 dB, but moved only one of four inputs below the unchanged detector threshold.

Replacing the complete tile was qualitatively different. At replacement strength 0.95, all four fixed-carrier scores moved below the frozen 0.17357069773071196 threshold, from source scores 0.7322, 0.6503, 0.3300, and 0.4161 to 0.097, 0.031, 0.158, and 0.087. Median PSNR was 53.85 dB, minimum PSNR was 52.56 dB, and minimum grayscale SSIM was 0.99811. All four paired score changes had the same direction, but four parents give a two-sided sign-test result of only p = 0.125. The probe was designed after the carrier family was known and is a causal local-score intervention, not an independent detector or removal validation.

A preregistered provider check stopped at its source control. Gemini returned the expected Google C2PA record but its pixel-verification stage remained at Connecting to Verify AI; no candidate was submitted after the unhealthy control. The next healthy oracle window must use the unchanged source control, the frozen 0.95 replacement, and a norm-matched sham. Until that paired batch returns a source-positive and candidate-negative pixel verdict, random-codeword replacement remains a stronger mechanism hypothesis rather than a SynthID remover.

A follow-up radial ablation separated the replacement tile into full-image frequencies measured in cycles per pixel. On the same four public positives, randomizing only 0.125-0.40 reduced the fixed score by a median 0.2093 at 56.48 dB median PSNR. The 0-0.125 band had no median effect, and frequencies above 0.40 changed the median score by only -0.0086. Lower-middle and upper-middle bands each contributed, with median changes of -0.0594 and -0.0508. This locates the measured Gemini carrier's most efficient causal surface in the middle band, not at the highest frequencies.

The ablation still operates on a folded periodic tile, so every edited coefficient remains a comb harmonic. It does not distinguish a sparse set of carrier lines from broadband coded noise between those lines. The next mechanism test must use continuous annular bands in local patch spectra, compare spatially independent and cross-patch-coherent perturbations at matched OKLab distance, and preserve a fixed random seed before any provider verdict.

Decision record

The program has four possible honest outcomes per provider:

Outcome	Product consequence
Causal signal and detector both generalize	Continue to pixel-only removal
Detector works but causal attribution fails	Ship no SynthID detector claim; retain as provenance research
Detector generalizes but pixel-only removal does not transfer	Keep local detection, retain regeneration fallback
Pixel-only removal clears the oracle with quality gates	Productize provider-specific detector and remover

Stopping at a failed gate is a result. It prevents a local surrogate, export fingerprint, or quality metric from being mistaken for control over SynthID.

Decision taken 2026-08-16

The second row is the outcome that occurred, on the evidence in this log. No local expert in this program detects SynthID. large-v1 detects the generation pipeline: its statistic is destroyed by a seven-pixel crop, while the published evaluation of the mark retains 99.97% TPR under aggressive crop and resize. The direction agreed is therefore to keep what works under its true name and to build toward provenance rather than toward a watermark claim.

Ship what is real, renamed. large-v1 identifies large Google-pipeline images with a measured 0/2,637 control rate and 8/11 on fresh out-of-distribution positives. It may be shipped as generator or pipeline identification, never as watermark detection, and its two failure modes belong in the product surface rather than in a footnote: it breaks on a crop that is not tile-aligned, and it breaks when the generation pipeline changes, as gemini-2.5-flash-image output already demonstrates at 0.063.

Develop toward provenance. Generator, pipeline and epoch identification is a task with abundant data, reliable labels and a measurable gate. U2 shows it is not solved, since the current features accept between a fifth and a half of foreign synthetic images, but it is the task the evidence actually supports.

Keep the watermark route open, blocked on access rather than on effort. A crop-robust detector of a keyed, perceptually masked mark needs a learned decoder trained on clean pairs. Google emits no unwatermarked image from any accessible path, and OpenAI's verifier may not be used to label a training set. If either constraint lifts, this route reopens; until then it is not an engineering problem.

One rule adopted from the failure. Every candidate local expert must survive a non-tile-aligned crop before it is calibrated. The test costs one line, it is the property the mark provably has, and its absence is why a pipeline artifact was carried for three months as a watermark detector.

Historical first milestone

The first local-research milestone was defined to produce evidence rather than shipping code. Its scope was:

1. the private-corpus schema and auditor;
2. an OpenAI authorization decision for use of the remote provenance verifier;
3. an independently verified status for candidate causal pairs;
4. a canonicalized OpenAI pilot set with hard negatives;
5. the D1 confound report;
6. the D2 low-texture carrier report with leave-one-group-out results;
7. a go or no-go decision for real-image detector training.

The completed experiments produced a no-go decision for a local universal OpenAI detector. The later official remote backend is a separate production route and does not retroactively turn provenance-labeled exports into pixel oracle training data.

Primary sources

• OpenAI, Content provenance.
• OpenAI, ChatGPT Images 2.0 system card.
• Google, Verify AI-generated images, videos, and audio.
• Gowal et al., SynthID-Image: Image watermarking at internet scale.
• Meta, VideoSeal.
• Meta, Watermark Anything.