`skills/remove-ai-watermarks/` teaches a coding agent to drive the CLI on
content the user generated, and refuses stock-agency and other third-party
paid-asset marks. It ships with this repository rather than as a fifth release
surface, and installs through `npx skills add` in any harness or as a Claude
plugin from `.claude-plugin/marketplace.json`.
`scripts/probe.py` measures capability rather than presence: a found binary said
nothing about whether the build can run a visible-mark command, so it writes a
blank PNG with zlib and struct and runs the installed `visible` and `invisible`
on it. Exit 2 is a state only a working stack reaches, and `invisible` checks
its extra before detection, so a signal-free image answers in 0.23s with no
model download. That distinction matters: without CUDA, `all` writes nothing
with the invisible extra installed and writes a useful visible+metadata result
without it, and both exit 1.
The skill also carries what a weak model gets wrong. It must report only what a
command in the session actually did, because a run on a machine with no pixel
stack produced a finished-looking removal report, complete with a size table for
a file nobody wrote. It must not read exit 2 as a verdict, because Click's usage
errors share that code. And it must get explicit consent before
`verify-openai-synthid`, which uploads the user's image to a third party.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A DEFINITELY verdict now runs a linear receipt gate on the CLIP-L-ft
vector Model 1 already computes: a hit publishes unknown and skips the
124-d provider pass. Trained on CORD-v2 train (800, CC BY 4.0) plus 200
synthetic capture-style receipts against 2,400 ai_train negatives; the
threshold is certified on CORD validation and a synthetic holdout with
no field-receipt pixels. Field receipts false-AI 19/59 -> 2/59, CORD
89/99 -> 0/99, ai_test DEFINITELY cost 7/1,847.
The Hub model and the private training repository were both called
raiw-models, so the training repo's README had to spend a paragraph telling
the two apart. Renamed on both sides: the Hub model to raiw-photo-classify
and the GitHub repository to raiw-classifier-training.
Done now because it is free now. classify.py landed after v0.36.0 and is in
no release, and the Hub model had zero all-time downloads, so no installed
package and no user carries the old id. Hugging Face redirects the old URL
and preserves counts, and the pinned freeze revision c0ac82b6 resolves on
the new name -- verified by downloading through the library's own constants.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* Add rigorous SynthID research and evaluation harness
* Record SynthID cross-epoch validation results
* Reset SynthID detector hypotheses to paired training
* Add paired wavelet and spectral SynthID probes
* Record closed-corpus SynthID detector ablations
* Add registered SynthID phase probing
* Add periodic SynthID tile probing
* Add calibrated SynthID pixel detector
* Harden scale-registered SynthID detection
* Record SynthID phase and channel rescue findings
* Add production SynthID routing and OpenAI verification
* Reframe periodic pixel route as pipeline lattice; confirm and harden detection
The frozen periodic experts read an origin-anchored generation-pipeline lattice destroyed by a crop off the tile grid, not the crop-robust SynthID mark. Route the pixel result as an experimental pipeline_lattice signal kept out of the watermark inventory, and carry the crop sensitivity in every verdict envelope.
Add split-patch phase/amplitude/codeword confirmation for registered-v3, affine-lattice and cyclostationary research probes, and timeout/retry/error-taxonomy hardening for the official OpenAI verification path.
* Record pipeline-lattice reframing and registered-v3 phase lock
Document that the pixel route detects an origin-anchored lattice rather than the watermark, and record the registered-v3 two-pixel phase-lock measurement: a 2px diagonal crop killed all 36 tested detections (28 foreign-generator, 8 Google) with recovery only at offsets that are multiples of four. Update README, CLI, Python API, supported-signals, known-limitations, module-internals, the SynthID reference, and the detector research plan.
* Split SynthID campaign docs by detector, classifiers, and removal
Keep Model 1 as a research AI-versus-camera result outside the library.
Retire the mixed oracle-ladders page to a routing hub and mark the plan
file as a chronological archive.
* Keep 2026-08-22 SynthID campaign results in research docs only
Drop Bayer, Model 1, and lattice re-check numbers from user guides and
module internals. The library surface is unchanged.
* Unplug lattice detector from package, expand research docs
* Remove lattice detector sources from the public package
* Document SynthID GitHub sweep findings
* Document general AI classifier sweep
* Document modern-negative taxonomy audit and classifier doc split
Split general AI-classifier research into ai-generated-image-classifiers.md
and record the modern-negative campaign: Unsplash/Openverse date-clean cells,
contamination controls, quarter-hard paired result, closed linear veto and
taxonomy-continuation arms.
pre-commit: 1) maintain.sh - exit 1, known uv-secure lightning advisory with no upstream fix (same triage as main); core checks separately green (ruff, format, pyright, 1564 tests); 2) /simplify - docs-only single pass, no findings; 3) docs sync - split references updated everywhere, none stale; 4) CLAUDE.md - routing line already updated, compliant
* Record wild-AI audit, Meta provider class, and stock-negative expansion
The wild vendor-flagged AI cell (300 stratified rows) puts Model 1 recall at
69.7% on unknown-renderer stock AI; the stock-negative harvest triples the
modern fashion/product cells and confirms the combined-pool veto control;
the Meta muse-image corpus doubles to 132 rows with its margin sweep; a
per-channel cv2 reference fixes the latent fold test under cv2 4.10.0.
pre-commit: 1) maintain.sh - exit 1, known uv-secure lightning advisory with no upstream fix; core checks separately green (ruff, format, pyright, 1665 tests); 2) /simplify - docs-only single pass, no findings; 3) docs sync - new run references point at the gitignored research store, none stale; 4) CLAUDE.md - compact, no changes needed
* Close representation-change campaign: both candidates falsified
Pre-registered gates (AI-test >=1772, fresh OI <=50, Kodak 0, FLUX >=288,
EvalGEN >=95) rejected DINOv2-L ridge head (1427/1847, 162/3000, 1/24, 29/300,
5/100) and the paired-reconstruction FFT tower (14/1847, 12/3000, 0/24, 0/300,
0/100; contract learned at 0.949 pairwise separation but no transfer to
production generators). Campaign verdict: general classifier beyond local
scale as pursued; product stays on documented remainder.
pre-commit: 1) maintain.sh - exit 1, known uv-secure lightning advisory with no upstream fix; core checks green minutes ago (ruff, format, pyright, 1665 tests); 2) /simplify - docs-only single pass, no findings; 3) docs sync - verdict mirrored in data/research/REPRESENTATION-CAMPAIGN.md, no stale refs; 4) CLAUDE.md - no changes
* Record full failure map, H2 router falsification, and H3 process detector closure
20 measured negative domains with batch-2 (memes, whiteboards, tickets, book
pages). H2 domain router falsified: 97.9% of AI-test looks structured, every
family 94-100%. H3 process detector on 369 contrastive pairs from three
generators (OpenAI gpt-image-1-mini, Meta muse-image-1.0, Gemini flash-image)
achieves 0.909 pairwise AUC on training pairs but transfers at AUC 0.55-0.60
to production AI; human receipts score higher than any AI family. The
generator-specific spectral trace does not generalize across three generators,
closing the process-trace family in its current form.
pre-commit: 1) maintain.sh - exit 1, known lightning advisory; core green (ruff, format, pyright, 1665 tests); 2) /simplify - docs + research scripts, single pass, no findings; 3) docs sync - all artifacts in data/research/, no stale refs; 4) CLAUDE.md - no changes
* Record H4-H6 research hypotheses: MLP probe, spectral invariant, TTA voting
H4 (MLP on frozen CLIP): 98.6% AI-test recall but 15.8% photo FPR — same
trade-off curve as quarter-hard. H5 (spectral invariant from CVPR 2025):
2.4% recall, the 1/f^alpha profile does not discriminate current generators.
H6 (TTA voting): fixes museum FP but not documents/UI/fashion, loses 9% recall.
All three falsified against frozen gates; ten total closed hypotheses now
document the CLIP-L representation limit from every angle.
Also records the provider cascade across all 20 domains (69.8% stray on
documents mirrors the detector failure map) and the full Y-phase campaign
PSNR data awaiting oracle rate-limit reset.
pre-commit: 1) maintain.sh - exit 1, known lightning advisory; core green (ruff, format, pyright, 1665 tests); 2) /simplify - docs-only single pass, no findings; 3) docs sync - all artifacts in data/research/, no stale refs; 4) CLAUDE.md - no changes
* Fix Y-phase frequency convention, record H4-H6 and campaign results
The initial Y-phase campaign targeted FFT bins 16-32 instead of spatial
periods 16-32 px (1/32 to 1/16 cycles/px). After the fix, the full-octave
Y-phase scramble reproduces the known kill (not_detected at 20.5 dB on s1
at native 1122x1402). Partial sub-bands are visually invisible (42.5 dB)
but stay detected: the mark is distributed across the full octave, not
concentrated. Point intervention is not viable within this frequency band.
H4 MLP probe: 98.6% recall, 15.8% FPR (same trade-off wall).
H5 spectral invariant: 2.4% recall (1/f^alpha does not discriminate).
H6 TTA voting: fixes museum FP but not documents/UI/fashion.
pre-commit: 1) maintain.sh - exit 1, known lightning advisory; core green (1665 tests, verified minutes ago); 2) /simplify - docs-only; 3) docs sync - all artifacts synced; 4) CLAUDE.md - no changes
* Add TC260 fifth class and per-class margin calibration to provider cascade
Five-head cascade (openai/google/tc260/meta/no_ai): openai 68.8%, google
76.6%, tc260 63.8%, meta 71.8% mean recall at uniform margin 0.50.
Per-class margins calibrated against 1.67% photo budget: openai 0.31,
google 0.29, tc260 0.39, meta 0.55. TC260 test recall 71.4% at calibrated
margins on 196 held-out images. Chinese-AIGC is a viable pixel class.
pre-commit: 1) maintain.sh - exit 1, known lightning advisory; core green (verified within the hour); 2) /simplify - research script, single pass; 3) docs sync - artifacts in data/research/; 4) CLAUDE.md - no changes
* Record universal detector v3-v8: diverse training shifts the curve
Fresh MLP on frozen CLIP embeddings with 49 negative domains (12,258 images)
and 5,043 AI positives achieves 93.0% AI-test recall with 39/49 domains at
zero false positives. Digital art FP dropped from 99.3% (Model 1) to 24.8%,
fashion from 62-71% to 12%, UI from 99.3% to 12%. Photo FPR 0.7%. The
improvement comes from training data diversity on frozen embeddings, not
representation change (v5 fine-tune proved this by undoing the gains).
Also adds .local-eval to ruff exclude so the gate covers only tracked files.
pre-commit: 1) maintain.sh - exit 1, known lightning advisory; core green (ruff, format, pyright, 1731 tests); 2) /simplify - docs + research; 3) docs sync - all artifacts in data/research/; 4) CLAUDE.md - no changes
* Add three-tier ensemble detector: DEFINITELY/POSSIBLY/LIKELY_HUMAN
Combines Model 1 and v8 into calibrated confidence tiers. DEFINITELY_AI
(both models agree) achieves 79.6% recall at 0.50% photo FPR with 46/49
domains at zero FP. The union reaches 93.3% recall at 1.80% photo FPR
with Kodak 0/24. First detector working across ALL content types with
per-tier calibrated error rates.
pre-commit: 1) maintain.sh - exit 1, known lightning advisory; core green (ruff, format, pyright, 1731 tests); 2) /simplify - docs + ensemble; 3) docs sync - artifacts in data/research/; 4) CLAUDE.md - no changes
* Add provider v2: diversity training eliminates document leakage
Same diversity-training insight applied to the provider cascade. Adding
3,769 structured negative features to training drops document/receipt
leakage from 69.8% to 5.1%, UI from 10% to 4.3%, fashion from 5-15% to
0-3%. AI recall improves: OpenAI 84% (was 75%), Google 88.8% (was 79%).
Photo leakage 1.4%. Worst regression: memes 47.9% (was 27%).
pre-commit: 1) maintain.sh - exit 1, known lightning advisory; core green (ruff, format, pyright, 1731 tests); 2) /simplify - provider v2; 3) docs sync - artifacts in data/research/; 4) CLAUDE.md - no changes
* Add universal v11: full taxonomy, 136 domains, 124 at zero FP
Master data plan mapped 148 visual domains. Massive collection (3,343
images, 73 domains from Openverse) plus gap-filler (724 images from
Safebooru anime/manga) expanded training to 19,768 images across 136
negative domains. Universal v11 (MLP-512 + Model 1 ensemble):
DEFINITELY 90.3% recall at 0.50% photo FPR, ANY 95.6% recall,
124/134 domains at exactly 0% FP, worst FP 13.8% (digital art).
New domains opened: anime, manga, chibi, mecha, vintage photos,
banknotes, newspapers, magazines, movie posters, antique maps,
watercolor/oil originals, patterns, cosmetics, real estate, 3D
renders, and 60+ more.
pre-commit: 1) maintain.sh - exit 1, known lightning advisory; core green (ruff, format, pyright, 1731 tests); 2) /simplify - docs + research; 3) docs sync - artifacts in data/research/; 4) CLAUDE.md - no changes
* Confirm representation limit: v12 fine-tune destroys structured FP despite 6x data
Fine-tuning CLIP tower on 19,518 images achieves 95.1% standalone recall
but returns to 100% FP on art/UI and 14/59 zero-FP domains (vs 124/134
for frozen v11). This is a representation property, not a data issue.
Frozen CLIP + MLP (v11) is the only working architecture. Also records
provider v3 trained on all negative domains.
pre-commit: 1) maintain.sh - exit 1, known lightning advisory; core green (ruff, format, pyright, 1731 tests); 2) /simplify - docs + research; 3) docs sync - artifacts in data/research/; 4) CLAUDE.md - no changes
* Document v13 plateau: 128/136 zero FP, frozen-CLIP ceiling confirmed
v13 with top-plan expansion (1,169 new images) improves domain coverage
to 128/136 at zero FP and reduces fashion FP to 5.3%, but DEFINITELY
recall plateaus at 89.6% and art FP at 15.9%. The frozen CLIP + MLP
architecture has reached its ceiling.
pre-commit: 1) maintain.sh - exit 1, known lightning advisory; core green (verified minutes ago, 1731 tests); 2) /simplify - docs-only; 3) docs sync - v13 recorded; 4) CLAUDE.md - no changes
* Document the 2026-08-31 photo-classifier freeze.
Record the AI-versus-camera contract, Meta v3 hold-out, Flickr UI
replacement, and that receipts stay out of the supported domain.
* Scrub Spaces catalog details from classifier research docs
* Record receipt layout-gate probe for the universal detector
The head was trained on mixed Chinese AIGC producers that share the
TC260 mark. openai and google stay provider classes. muse-image stays
a model class.
openai, google, and tc260 stay provider classes. The Muse Image head
is public as muse-image, not meta. The freeze file still keys that
head meta_muse_image.
Ship Model 1 then gated Model 2 as an explicit CLI and library call.
identify, has_invisible_target, and all do not import it. Weights stay
on Hugging Face wiltodelta/raiw-models at a pinned revision.
The four-fixture Google calibration showed a clean YuNet face-count split:
both zero-face text cards need 0.25 first-clean, both face fixtures clear
at 0.12 with zero cross-source spread. The adaptive arm routes face
content (face_count > 0) to the measured 0.125 rung, saving 6.4 dB PSNR
and doubling face identity cosine vs the flat 0.40 floor. The split is
Google-SynthID-specific: OpenAI's face fixture was harder than its text
fixture, so no other cohort gets an adaptive arm.
Extract TwoStageZImagePipeline shared base from QwenZImagePipeline so all three
profiles are siblings, not a chain. Add ChromaZImagePipeline on that base with
diffusers ChromaImg2ImgPipeline bf16, neutral calibrated prompt, guidance 5.0,
four effective steps with Diffusers truncation compensation, no Canny.
Flat vendor floors from the 2026-08-29/30 oracle calibration across OpenAI
(0.09), Microsoft InvisMark (0.125), Google (0.40), Meta Content Seal (0.17),
seed-stability verified on Meta, full-path face-stage interaction clean through
the Gemini oracle. Research record in docs/chroma1-engine-research.md with the
matched-strength addendum showing Chroma1 regenerates better per unit strength
but its boundaries scatter wider than qwen's. Fix text-manifest guard to reject
chroma-zimage at the boundary, update all profile references from 'both' to
'all three' across src, docs, README, and python-api.
Jimeng CN video wordmark: no separation (1/5 positives vs 2/25 negatives hold
a stable 0.35 run). RunningHub video: separates but the cohort is two
positives, below the registration bar. The top-left pill on video is parked
by analysis: its false-fire mode is a static textured corner, which temporal
recurrence cannot screen.
detect_doubao_frame reuses the image engine's synthetic doubao alpha as its
video template, so the glyph cannot drift between registries. The policy is
seedance-style single-floor, calibrated by sequential decode over 39
TC260-confirmed Doubao corpus videos and 25 negatives: a 12-frame stable run
at 0.35 accepts 33/39 positives and 0/25 negatives (0.30 already accepted a
negative), and through the shipped arbiter identify_video selects 32/39 with
no false selection. Also documents the measured impossibility of a bare-pill
removal arm with the current detector (68 OCR-confirmed pills max 0.317 vs
clean negatives max 0.353 -- no separating threshold exists).
The OCR badge census split the 343 Microsoft-C2PA uploads into 86 badge
carriers and 257 true badge-less files (the watermark is a per-user opt-in).
Badge-less scores max 0.251 / p99 0.213, so the 0.266 relaxed gate admits the
three faint badges in [0.251, 0.38) with zero measured false fills; re-verified
on the provenance path as 3 band detections and 0/257 badge-less fires.
Tiled diffusion was never provider-oracle calibrated with verified text
restoration: the tiled VAE donor path ran anyway and produced results no
oracle had certified. The combination is now rejected at both the
pipeline and the engine seam (ValueError with the reason), and the CLI
help no longer implies support. The invisible help is generalized and
the metadata container list corrected (MKA/OGA/Opus/AAC).
scripts/contentseal_transforms.py reproduces the deterministic crop,
resize, and JPEG variants of the Content Seal corpus from manifest.csv,
hash-verifying every output; its README gains scripts/README.md context
and new data tests. The corpus README is honest about the one crop the
daily oracle limit left unchecked, and the eval CSVs carry the updated
verdicts. The byte-scan SynthID suppression hoists its soft-binding
lookup so the guard is computed once.
Staged on top of 0.33.1; no version bump in this commit.
Full Meta Muse Image support in the invisible-removal path:
- QWEN_ZIMAGE_META_STRENGTH = 0.1: derived by the standard
worst-boundary-plus-cross-source-spread method over five oracle-bracketed
generations (data/contentseal/manifest.csv)
- Auto mode: vendor_for_strength routes a file whose only provenance is the
standalone AI IPTC trainedAlgorithmicMedia tag onto the meta cohort; C2PA
issuers win first, so Google/OpenAI/Microsoft routing is unchanged. Muse
WebP outputs place the XMP in a tail chunk, so the scan uses the shared
chunk-aware metadata.scan_head rather than a plain head read
- Explicit override: --vendor on invisible/all/batch and
InvisibleOptions.vendor name the cohort on stripped files; naming a cohort
asserts the watermark is present, so the no-signal gate treats it like
--force at both the CLI and API seams
- sdxl-zimage has no measured Meta rung: an explicit meta vendor falls to
the conservative unknown 0.25 rather than inventing one
- identify emits a Content Seal caveat pointing at the removal path
- The legacy visible 'Imagined with AI' mark stays unregistered: a dedicated
sample hunt (newsroom mockups, community posts, press screenshots, dead
imagine.meta.com, broken Wayback captures) found no pixel-verifiable
capture, and the registry rule forbids encoding a corner without one.
erase --region remains its removal path; outcome recorded in the landscape
Co-Authored-By: Claude Fable 4.5 <noreply@anthropic.com>
Muse Image (muse-image-1.0, Meta Superintelligence Labs, 2026-07-07) ships
every output with Content Seal, a proprietary invisible pixel watermark, and
no visible mark. Establish support documentation and a verified corpus:
- data/contentseal/: five own generations via the Meta Model API, every
oracle verdict recorded in manifest.csv (44 rows, settled-text protocol,
fresh-navigation variant for calibration rows)
- Oracle: meta.ai/identification web tool only; no API endpoint exists in
the Meta Model API (verified against dev.meta.ai/docs); internal REST
pair documented with its server-side sliding-window rate limit
- Removal: default qwen-zimage profile clears Content Seal (oracle-verified
on the worst source); strength floor derived at 0.1 by the standard
worst-boundary-plus-cross-source-spread method, recorded but not encoded
as a constant since no provenance signal routes Muse output onto a vendor
cohort
- Seal robustness measured: survives resize, JPEG q85, metadata stripping,
CDN WebP transcode; dies to center crops and diffusion regeneration
- tests/test_contentseal_corpus.py guards manifest integrity
Co-Authored-By: Claude Fable 4.5 <noreply@anthropic.com>
High-confidence C2PA attribution required signingCredential.trusted, a status code
the reader emits only when a trust anchor list is loaded. None ships, so from 0.27.0
through 0.30.0 the branch was unreachable in production for every vendor: an intact,
cryptographically bound manifest scored the same medium as a fallback parse that
validated nothing, which collapsed the one distinction the official reader exists to
draw. A hand-built info dict stamping that code kept the branch green in the suite.
Confidence now follows the binding. Signer trust and certificate expiry stay visible
as their own dimensions and as caveats, because a trust list that was never
configured is a missing input, not a finding against the credential. Every committed
provenance fixture with a reader result and an intact binding now reaches high
confidence, and test_no_committed_fixture_reports_a_trusted_signer guards the
reachability itself rather than a synthesized status set.
Revocation joins binding and signature failures as disqualifying. It arrives only on
signer_validity, so a check reading the other two returned a confident AI verdict off
a credential the issuer had disowned, with an empty integrity_clashes -- quieter than
a hash mismatch on the same file. Expiry stays non-disqualifying: it does not imply
the signed bytes changed, and a signature genuinely made outside validity already
arrives as claimSignature.outsideValidity.
The rule now lives in one place. _validation_fields maps status codes to the four
dimensions and names the failures that moved one; c2pa_info_has_invalid_credential
maps dimensions to disqualified. The ingredient-reachability walk and the
user-visible reason both consume that path instead of re-classifying raw codes, so
adding this one rule no longer means editing three layers in lockstep.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Reachable c2pa.soft-binding assertions now surface their exact alg and the
bounded printable block value next to the normalized vendor label; a value
without its algorithm stays hidden because it cannot be attributed.
com.microsoft.invismark.1 uses that value as the pixel-watermark identifier
in Microsoft Paint output, so identify names it, metadata carries it, and an
InvisMark soft binding keeps the invisible-removal gate fail-safe even after
C2PA asset binding goes invalid. Content fingerprints still do not trigger
pixel regeneration. Removal is verified against Microsoft's Content
Provenance Detection API, which reports Watermark and C2PA separately; the
protocol and the pixel-identical control requirement are documented.
Implemented in a parallel session; verified, gated, and committed by pi.
pre-commit: 1) maintain.sh - exit 1 on the known uv-secure lightning PYSEC-2026-3624 triage (no fix available, unchanged from 0.29.0); ruff, pyright src/, and 1391 tests passed separately; 2) /simplify - single-pass, clean; 3) docs sync - five docs updated by the author session, no remaining references found; 4) CLAUDE.md - invariants recorded in module-internals, no change needed
Each Haar pass is one flat pywt.downcoef call over a raveled strip instead of
pywt.dwt(..., axis=1)[0], and the plane is walked in strips so no full-plane
float64 intermediate exists. Exact only while the last axis is even, so
_approximation raises on an odd width rather than returning wrong bits, and
TestRaveledHaarPass pins both that raise and the downcoef/dwt equivalence a
pywt upgrade could take away.
Drops the block constructor knob: the fold chains are written for 4, nothing
ever passed another value, and a knob that silently decodes wrong is worse
than no knob.
Peak RSS 111 MB to 21 MB on a 4.3 MP image; the decoder itself 0.011s to
0.007s, which is only 0.4% of identify() now that it is under 2% of the run.
Output bits and detector verdicts over 200 sampled data/ images, two
synthesized carriers and eight degenerate shapes are byte-identical to the
pre-vectorization decoder.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Output stays bit-identical: decoder bits and detector verdicts recorded over
200 sampled data/ images plus two synthesized carriers before and after, and
the record is byte-identical.
Measured on a 1536x2816 image -- decoder 0.280s to 0.016s, warm identify()
1.757s to 1.365s with both arms timed in one process.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The probe shipped with no tests while every sibling research script has them,
and it needs neither a model nor a GPU, so nothing justified the gap. Its whole
output is a ranking and a mispaired comparison still prints a plausible number,
so the pairing tests carry more weight than the metric ones.
The centerpiece builds a correctly phased and a wrong-phased delivery of the
same clip. Both have six frames, so the frame-count check the probe used to rely
on passes for both; the misaligned one scores near 9 dB against 40+. A separate
assertion pins that the probe binds the engine's sampler rather than a copy,
which is the contract a same-phase private copy would otherwise satisfy
silently. This file now holds the only constraint on that sampler's phase.
CI ran none of this. The test matrix installs no ffmpeg, and the job that does
install it ran one class from another file, so five of these tests would have
skipped on every runner. The video-e2e job now collects this file too.
Recorded in .claude/rules/development.md: mutating the sampler's >= to > proves
nothing, because the 1e-9 epsilon makes both branches agree on exact equality.
The phase or the period is what has to move. That inert mutation passed a green
suite here before the real one failed five tests.
Dropped the downscale test's PSNR bound: solid colors survive both the downscale
and the upscale exactly, so any score there pins the local ffmpeg's chroma
rounding rather than the probe. Geometry is the assertion that means something.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The registries are raw substrings and the shortest tokens are four and five bytes
(`Bria`, `Adobe`, `Canva`). Over a megabyte of compressed pixel data such a sequence
turns up by chance: `Bria` matched inside the entropy-coded scan of 4 of 14,707
corpus JPEGs, in none of which the manifest names Bria. The rate is what a four-byte
pattern predicts on that corpus, and the Bria entry asserts AI, so a chance match can
declare an image AI-generated rather than merely mislabel its signer.
`_metadata_region` gives the registry scans the container's metadata: JPEG marker
segments before the coded scan, PNG chunks other than IDAT, both trailers, and
whatever `scan_head` appended past the window. Every other check keeps the full
buffer -- their markers are long and distinctive. A container that does not parse is
returned whole, since dropping real evidence to avoid a chance match is the wrong
trade. `c2pa_marker_in` already refuses a bare `c2pa` substring for this reason;
this is the same defence for the registries.
Verified the way the rules require for a change that MOVES a verdict: over all 48,905
corpus images, exactly one file changed, the one named in advance, from
"C2PA Content Credentials (Bria Artificial Intelligence)" to "(unknown signer)".
Record-path parity is 0 disagreements, down from 75 when this work started.
The audit's own baseline comparison is fixed here too. It compared confidence and
signals only, and so reported "0 changed" for the run whose single intended
correction was a watermark line -- the change it exists to show.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`scripts/ai_score.py` and the dataset scanner that fed it are gone: the detector
they trained is not something this project runs, and the corpus lived outside the
repository anyway. Nothing else referenced them.
The scanner's pixel layer was worth keeping, so it moves into the package as
`pixel_evidence.py` -- six families of scale-robust statistics (block-DCT histograms
and Benford deviation, FFT band energies and CFA peaks, high-pass residual, error
level, gradient, colour) measured in a single shared decode. The arithmetic was
verified against the scanner over 60 corpus images, families and artifacts alike,
before the scanner was removed; that comparison is no longer possible, which is why
the tests now pin behavior instead: determinism, empty-not-wrong on images too small
for a family, and one failing family not taking the others with it.
It has no consumer. Nothing in the package reads it, and the module says so.
`artifacts=True` returns the spatial layer -- perceptual hash, 128px thumbnail,
coarse ELA/residual/phase maps. Those identify the source image rather than describe
it, so they are opt-in and separate: everything else is a scalar or a fixed-length
histogram nothing can be reconstructed from.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A full-corpus audit of the record path against the file path found 75 of 48,905
images disagreeing, and 74 were one gap: the SynthID byte scan for containers whose
manifest no parser reaches lived in `get_ai_metadata`, an extractor the record path
does not run. The record silently reported no SynthID for images `identify` flagged.
Moving the scan into `identify_from_evidence` fixes it by construction rather than by
copying the rule into a second extractor -- the same shape `soft_binding` already
uses. Its byte checks mirror `metadata.synthid_source` literally instead of reusing
the broader `has_c2pa` / `c2pa_source_kind` derived above, so the file path's answers
do not move: verdicts over a 4,000-image sample are byte-identical.
`scripts/record_parity_audit.py` is the audit itself, now repeatable. It walks a
dataset, judges every image through both seams with the record round-tripped through
JSON, and reports disagreements by field and by signal. The rule in
`.claude/rules/development.md` says to re-run both sides of this seam after changing
either; this is what to run.
Both timing and audit scripts now put the package's OWN `src` on the path. From a
worktree an editable install resolves to the main checkout, so the audit imported a
different tree than the one under test -- the failure the same rules file warns about,
reproduced within an hour of writing it down.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Three gaps found while measuring the record path against the file path, each one
a signal the library could not see:
WebP stores `XMP ` after the pixels, so on any WebP above the scan window a fixed
read stops short of the label. `_riff_late_metadata` steps over the coded image to
reach it, the RIFF analogue of the existing PNG and ISOBMFF readers. Three corpus
files hid an IPTC "Made with AI" tag and a C2PA `trainedAlgorithmicMedia` there.
The decoder-backed fallback now covers only what it is actually for -- metadata the
raw bytes do not spell, such as a compressed PNG `zTXt` packet.
A C2PA reader failure returned the same `None` as a file with no manifest, so a
verdict could fall back to the raw byte scan with no trace anywhere. Failures now
log at warning and only genuine ones do: a file without credentials never reaches
that branch, and an unsupported container is demoted to debug through the reader's
own `C2paError.NotSupported`. The first corpus run with it found a truncated PNG.
`scan_dataset.py` never registered the pillow-heif opener it declares as a
dependency, so every HEIC was scanned as unreadable -- no EXIF, and a pixel layer
that was 397 of 406 features NaN instead of 136.
`_riff_late_metadata` caps its total like `isobmff.scan_c2pa_region` does. Clamping
each chunk to the bytes remaining is not enough on its own: one chunk can declare a
length spanning most of the file, and this runs on the memoized verdict path over
images from arbitrary sources.
Also lands `identify_metadata_record` and `ProvenanceReport.to_dict()`, the
one-call entry point and the versioned JSON contract for the record path.
Record-vs-file equality holds over 3,478 corpus images, and the eight files these
fixes recovered still report AI.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`collect_metadata_record` returns a JSON-safe record carrying an image's
provenance metadata regions -- never its pixels -- and the existing
`evidence_from_metadata_record` + `identify_from_evidence` build the verdict
from it without opening the file. The contract is equality with
`identify(path, metadata only)`, verified over the tracked fixtures and over a
local corpus of 3,478 images (every file carrying a rare signal, plus a random
slice): zero differences.
Three placements defeated earlier drafts and each is now a rule with a test:
the `scan_head` buffer is the head CONCATENATED with late metadata, so a
structural walk must read the raw head instead; Samsung splits its evidence
between a post-EOI trailer and the coded scan; and PIL's info keys must be
emitted in the file path's candidate order, since the first token match wins.
Also fix a real detection gap found while establishing that equality: a label
the decoder can read but a raw byte scan cannot -- a compressed PNG `zTXt`
packet, or a WebP XMP chunk past the scan window -- was invisible to
`identify`. Eight corpus files carrying a China TC260 AIGC label or an IPTC
"Made with AI" tag were reported as no signal at all.
`scripts/detection_timing.py` and its report script measure the metadata path
per method; they write outside the repository and are read-only over a dataset.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The shipped profile was certified by one oracle row, but only noise_std was
pinned: long_side and fps -- two thirds of what the verifier was actually shown
-- could move with a green suite. The test now derives the pin from
data/evaluations/video-synthid-oracle.csv, so a default without a certifying row
fails.
The certified profile is a perturbation-to-signal ratio, not a bare noise_std.
sd-vae-ft-mse publishes no scaling_factor key, so 0.18215 comes from the
AutoencoderKL class default under an upper-unbounded diffusers pin. The loader
now gates that value, carries it on VideoVaeRuntime, and passes it into encode
and decode so the validated value is the applied value. video_synthid_sweep.py
loads through the same function: the harness producing the certified rows was
the one path exempt from the gate it exists to feed.
psnr_db is measured against the already-resized frame and before the encoder, so
it cannot see the downscale, the decimation, or the codec, and no in-loop metric
can. scripts/video_fidelity_probe.py scores the delivered file end to end,
streaming the way the engine does and sharing its frame-selection rule rather
than copying it -- a frame-count check cannot catch a rule that reorders frames
without changing how many.
The manifest gains source geometry, vae, track, verbatim verdict and session
fields. The two 2026-07-31 rows keep them empty: they were never recorded and
are not recoverable. Verdicts now have four states, because the verifier's
unclear reading logged as not_detected is the silent regression the manifest
exists to prevent.
docs/video-synthid-quality-research.md records the research behind this: the
noise axis is worth about 2 dB and is nearly exhausted, resolution is the real
prize but is an uncertified destruction axis rather than a free win, and every
proposed autoencoder swap was refuted. First local measurements included.
Verified: engine output is byte-identical before and after the refactor on a
locally built clip, at noise_std 0.00 and 0.15.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Victor Kuznetsov