remove-ai-watermarks/docs/verification-plan.md

Full verification plan

Verification log and historical plan. Dated results describe the code and datasets at the time of each run. They are evidence for maintainers, not a current CLI contract. Recheck open items against the current code and issue tracker before treating them as work that remains.

How we convince ourselves the library actually works, across its whole surface, on real data.

The 2026-09-08 code and documentation audit records current reproducible defects, documentation discrepancies, review coverage, and the checks needed to close them.

This is the pre-release and periodic-audit plan. It is deliberately organized by oracle strength rather than by module, because the hard part is never "call the function" -- it is "know what the right answer was". A sweep with no oracle proves only that nothing threw.

Performance depends on format, enabled extras, and hardware. Measure it locally with representative inputs; do not publish private dataset sizes or run statistics.

Public test sources

source	role
data/synthid/	labeled SynthID oracle fixtures and evaluation manifest
data/fixtures/provenance/	deterministic format and provenance fixtures
data/calibration/<vendor>/	minimal controlled inputs used to derive detection silhouettes
synthesized inputs	constructed ground truth for pure and format-level checks

Tier A -- self-evident oracles (full corpus, unattended)

Properties that are true or false without anyone labeling anything. These are the backbone: they scale to large local datasets and catch regressions with zero human cost.

A1. Recorded-verdict regression

A dated local cache of identify verdicts can be replayed against the same inputs to detect drift in verdict, platform, confidence, or signal set.

Caveat that makes this honest: a diff is not automatically a bug -- the sidecars were written by older versions, so intended improvements also show up. The output is therefore a classified diff (new detections / lost detections / changed platform / changed confidence), reviewed once, then re-baselined. Lost detections are the alarm.

Implemented as scripts/sidecar_regression.py (resumable, ~1.5 h at 8 workers).

A 2026-08-15 metadata-only C2PA regression audit over the local historical corpus caught two structured-parser gaps before release: an exact Firefly claim generator without a repeated source type, and a Dreamina generator carried by a reachable ingredient under a generic update manifest. Both now have focused tests. The same audit confirmed that invalid hash/signature claims lose origin attribution without losing the C2PA inventory, and exposed the absence of production trust anchors in the SDK defaults. Dataset-derived counts and identifiers remain in the gitignored audit output.

Local run protocol

Re-run identify against locally recorded sidecars, classify losses separately from intended new detections, and keep generated reports under .local-eval/. Do not commit dataset-derived counts or identifiers.

A2. Parity: whatever we detect, we must be able to remove

For every image where a signal fires: remove, re-scan with the same oracle, assert quiet.

• metadata: scripts/metadata_removal_audit.py;
• visible: scripts/visible_removal_audit.py, once per backend.

Detection does not depend on the fill backend. Run scripts/visible_positives.py once and pass its output through the audit's --paths-file option.

Local parity protocol

Run detection, removal, and re-detection over a representative local set. Confirm that decoded pixels remain unchanged for metadata-only operations. Keep reports untracked.

A3. Byte-level invariants

• no-op remove_visible returns the ORIGINAL bytes (not a re-encode)
• pixels outside the fill mask are bit-identical to the input
• JPEG metadata strip is pixel-lossless on the DEFAULT path (--remove-all re-encodes by design -- see metadata.py; assert the split, not losslessness everywhere)
• lossless source formats survive a misnamed extension

A4. Idempotence and order-independence

• remove_visible(remove_visible(x)) == remove_visible(x)
• strip(remove(x)) == remove(strip(x)) in signal terms
• a second identify on a cleaned output reports no metadata signals

A5. Contract sweep across every parameter choice

scripts/smoke_matrix.py covers every choice-valued flag on fixtures. Its knob rows were rewritten when the CPU/MPS profiles and the ESRGAN chain were removed: most now assert a knob is REJECTED, and the accepted-knob rows skip without a CUDA device, so the row count is host-dependent rather than the fixed 68 recorded here before. Extend from fixtures to a stratified corpus slice (~500 images spanning format x provenance x aspect ratio), asserting exit-code semantics rather than just absence of crash.

Known trap to encode: exit 2 is triply overloaded (no-visible-mark, no-invisible-signal, Click usage error). A wrapper cannot distinguish them without parsing stderr. Either the sweep asserts on stderr, or the codes get split -- the latter is the better fix.

Coverage method

Compare the flags and values exercised by the matrix against the options declared by the CLI. Include optional backends, batch modes, tiling, and the ffmpeg audio/video strip. (Region-targeted diffusion composition was on this list until the parameter was deleted: nothing but a caller-less convenience wrapper ever reached it, and the user-facing region path is erase, which inpaints instead.) The gap to find is not only "logic untested" but "never executed on real data", which is precisely what this campaign is for.

Bug found by the extension: --steps below ~7 crashed inside torch

Fixed by deletion. --steps no longer exists, on the CLI or in the Python API, so this class of failure is unreachable. Kept as a record of why.

Effective timesteps were int(steps * strength). At the vendor-adaptive default strength (0.15, or 0.10 for OpenAI) any --steps under 7 rounded to zero, and the pipeline died with a raw traceback:

RuntimeError: cannot reshape tensor of 0 elements into shape [0, -1, 1, 512]

Fully valid CLI arguments, no special flags, no --force. The value was accepted, the crash was a torch internal, and nothing told the user that steps and strength interact. The considered fixes were a clamp to >=1 effective step or an up-front validation naming both values; what shipped instead is that each stage owns its own distilled schedule and no caller can set it. The general lesson stands: a knob whose valid range depends on another knob's value needs the interaction validated where both are known, or it needs to not be a knob.

Method note: the first run of the knob rows failed 12 times with this identical error, which read like twelve broken features. It was one bad harness parameter (--steps 4) sitting on top of one real bug. An error that is IDENTICAL across unrelated rows is evidence of a common cause, not of many faults -- check the shared input first.

Tier B -- constructed ground truth (automatable, no labeling)

Where reality gives no answer key, build one. This is the tier that closes the two biggest holes: fill quality and detector response at the edge of the operating range.

B1. Fill quality with a true reference

Real marks have no clean counterpart, so quality has only ever been eyeballed. Construct it instead: take a clean corpus image, stamp a known mark at a known position (the captured alpha maps make this exact), remove it, and compare against the true original.

Yields PSNR / SSIM per --backend (cv2 / migan / lama), sliced by background class, which is exactly the axis where the docs say quality varies but no number exists.

Implemented as scripts/fill_quality.py. Two reporting rules are load-bearing: score INSIDE the footprint (whole-frame PSNR sits near 60 dB whatever the backend does), and use the MEDIAN (a fill that reproduces a flat background exactly scores PSNR=inf, and one inf makes a mean inf -- the first run reported "+inf" for every flat bucket).

tests/test_visible_fill_quality.py is the bounded CI smoke layer: every registered image mark is constructed on two deterministic backgrounds and sizes, then checked for stamp coverage, one-fill detector clearance, outside-mask pixel equality, and conservative regional PSNR/SSIM floors with the cheap OpenCV backend. The full local harness keeps the backend comparison and texture slicing out of CI, where model downloads and a large clean corpus would make the gate expensive and nondeterministic.

Local fill-quality protocol

Construct marked images from clean local references, compare each backend against the known original inside the affected footprint, and keep the report under .local-eval/.

Public clean-corpus run, 2026-09-04

The local source pool combined 186 Unsplash Lite photos and 235 Openverse Flickr Creative Commons images. All 421 manifest entries were unique, every referenced binary existed, and every SHA-256 matched before sampling. The report records only source filenames and measurements under .local-eval/; no image bytes or private corpus facts enter the repository.

The 60-source, three-backend run produced 1,980 valid measurements for the 11 non-Samsung marks. Every mark/backend pair had positive median PSNR recovery over the stamped input. In paired comparisons, LaMa exceeded OpenCV by a median 1.20 dB on 610 finite pairs and moved in the favorable direction on 432 versus 175 pairs (two-sided sign-test p=4.86e-26). MI-GAN exceeded OpenCV by 0.31 dB in PSNR, but its median SSIM was 0.0136 lower; backend preference is therefore metric- and content-dependent rather than a universal ordering.

The first Samsung rows from that run are rejected evidence: the forward model multiplied the already-solved alpha asset by 0.38 a second time and composited toward the generic 238 luma rather than the measured pure-white glyph. After fixing that construction, a factorial ablation changed only the removal mask on the same 60 sources. The detector-aligned sparse mask improved PSNR over the old rectangle on 55/60 OpenCV pairs, 55/60 MI-GAN pairs, and 57/60 LaMa pairs, with median gains of 14.98, 10.38, and 14.87 dB (p=1.04e-11, 1.04e-11, and 6.25e-14). Relative to the correctly stamped damaged input, median recovery was +13.27 dB for OpenCV, +9.31 dB for MI-GAN, and +15.49 dB for LaMa. These are constructed-reference mask quality results, not production recall.

B2. Detector response curves

Use scripts/detector_response.py to sweep mark size, opacity, background, and aspect. Report detection and maskability separately. Generated reports stay under .local-eval/.

The corrected 2026-09-04 run contains 28,800 stamped cells and 1,800 matched clean controls over 150 public clean backgrounds. Every mark had 0/150 control fires (per-mark 95% Wilson upper bound 2.5%), and every detected cell was maskable. Nominal constructed detection ranged from 54.0% for the Jimeng pill to 96.7% for Yuanbao among the non-Samsung marks. Samsung reached 20.0%; the earlier 0.7% value is invalid because it came from the double-attenuated renderer. Samsung still falls sharply on busy synthetic backgrounds, so this grid identifies a detector-recall follow-up; it does not justify lowering the threshold because the grid is adversarial and is not a production-recall sample. The corrected report is .local-eval/detector-response-corrected-n150.jsonl.

The 2026-09-05 Samsung follow-up changed the front end, not the threshold. Continuous top-hat with an exact (1.0,) ladder scored the three retained real positives at 0.82-0.90 and produced 0/421 fires on the full public clean pool (maximum 0.38; 95% Wilson upper bound 0.9%). The provenance factor is 1.0: the inherited 0.70 relaxation would admit 14/421 controls at its 0.28 gate and recovered no retained real positive. On correctly stamped nominal marks over those same controls, detection improved from 81/421 (19.2%) for binary to 321/421 (76.2%) for top-hat. All 321 boxes exactly matched the known stamp geometry and were maskable. After one OpenCV fill, 320/321 no longer detected; the remaining result is reported by post-removal validation rather than retried. The constructed rate is a controlled failure-mode measurement, not production recall.

The complete B2 harness was then rerun on the same deterministic 150-source selection into .local-eval/detector-response-samsung-tophat-n150.jsonl: 28,800 stamped cells plus 1,800 controls, with all 204 expected rows present for every source. Samsung's nominal cell improved from 30/150 (20.0%) to 120/150 (80.0%); all 120 were maskable, and controls remained 0/150 (maximum score 0.3543). Across the full adversarial grid, 399/2,400 Samsung cells detected and all 399 were maskable. The key sets matched the previous report, and all 28,050 non-Samsung rows were identical field for field, so the measured change is isolated to Samsung.

B3. Invisible round-trip, positive-control gated

The open DWT-DCT detector is positive-only and carrier-fragile: "not found" on a fragile carrier proves nothing (measured: chatgpt-1.png recovers 114/128, below the 118 gate). Every invisible assertion must first embed on the SAME carrier and confirm recovery, and degrade to a skip rather than a pass when the control fails. Already implemented in smoke_matrix.py; apply the same discipline anywhere else this detector is used.

B4. Resource ceilings

Peak RSS and wall time per backend x input size, up to 25 MP. The memory-constrained CPU tier is a real constraint (MI-GAN must stay ~0.6-0.9 GB by cropping around the mask); a regression here is invisible today and would only surface under load.

Tier C -- human-labeled accuracy (bounded by labeling effort)

The machinery exists: visible_recall_sample.py -> visible_sheets.py -> visible_groundtruth.py -> visible_eval.py.

• Recall needs representative sampling per mark and aspect ratio.
• Precision must be benchmarked before and after every detector change with --vs <snapshot>.
• Coverage is separate from tuning. A missing vendor detector cannot be repaired by changing another detector's threshold.

Three harness rules are load-bearing and must not be relaxed: score a mark only within its crop's adjudication scope; take provenance from metadata, never from labels; and never report recall from the detector-sampled set.

Tier D -- external oracles

The unified maintainer workflow, execution-slot schema, immutable batches, and result recorder are documented in provider-oracles.md and implemented by scripts/provider_oracles.py.

Proprietary watermark removal cannot be verified locally by design -- no public decoder exists. Each vendor has its own oracle and it covers only that vendor's content: the Content Provenance API first and openai.com/verify as an explicit fallback for OpenAI, the Gemini app for Google, and Microsoft's Content Provenance Detection API for InvisMark. The Microsoft API reports Watermark and C2PA separately. It therefore needs a pixel-identical metadata-stripped control: the control must lose C2PA while remaining Watermark-positive before a candidate's negative result can be attributed to pixel regeneration. A quiet metadata proxy is not proof the pixel watermark is gone.

OpenAI's API documentation says not to use repeated queries to reverse-engineer, remove, or evade a watermark. Using it as an adaptive research oracle therefore requires explicit authorization. Without that authorization it must not become a training loss, search loop, or automated removal gate. The provider-specific detector and pixel-only removal research protocol is in synthid-detector-removal-plan.md.

Scope honestly: this tier certifies strength floors on a handful of images per vendor, and that is all it can do. See docs/synthid.md.

D1. Sampling frame

Select a representative local sample without committing images, identifiers, dataset sizes, or oracle results. Stratify by signal family and preserve the sampling method outside the public repository.

D2. The oracle is the bottleneck, not the GPU

Measured on MPS (2026-07-19), single invocation of invisible:

--max-resolution	256	384	512	768	1024
wall time	37.9 s	58.9 s	59.4 s	65.2 s	118.3 s

384/512/768 are indistinguishable, so below ~768 the cost is dominated by fixed model load, not diffusion. Confirmed by batching: 4 images in one process took 105.5 s (26.4 s/image) against 59.4 s/image one at a time -- roughly 40 s fixed overhead per invocation and ~15 s marginal per image at 512 (rough: run-to-run variance is large).

Two consequences for the harness:

• Amortize the fixed cost: one long-lived process over many images, never one invocation per image. That is a 2-4x win. Shrinking below 512 is not.
• The binding constraint is the external oracle's throughput, which is manual and rate limited. So do not run a uniform grid; spend each oracle check where the answer is uncertain -- bisect strength per content class to certify a floor in ~10 checks instead of ~100.

D3. Mandatory control before trusting any reduced-size run

--max-resolution is downscale -> diffuse -> Lanczos upscale. If the resize round trip alone damages SynthID, the oracle goes quiet for a reason unrelated to removal and the result does not transfer to production at native resolution.

Before any reduced-size sweep, run the resize round trip with no diffusion and put the result through the vendor oracle. If the watermark survives, the reduced size is a valid test bed; if it does not, reduced-size results are measuring the resizer. This is the same failure shape as the imwatermark carrier-fragility trap in B3: an oracle that falls silent for the wrong reason reads exactly like success.

docs/synthid.md cites ~99.98% TPR across 30 transforms including resize, which predicts the control passes -- but that is Google's claim about their own decoder, not our measurement, so it is a hypothesis to test, not a reason to skip the control.

D4. Video candidates require a matched transcode control

Video experiments add frame sampling, resizing, frame-rate conversion, and a final video codec around the actual attack. scripts/video_synthid_sweep.py therefore emits control.mp4 from the same selected frames and encoder settings as every VAE candidate.

Verify the control first in a new Gemini chat by invoking @synthid and using the supported built-in content-verification question. Continue only when the provider oracle still detects SynthID in it. Verify each candidate in its own new chat. A generic response that discusses visual clues or metadata is not an oracle result. Do not ask the chat model to reinterpret or second-guess the built-in verdict; that follow-up is ordinary model reasoning. Record only the explicit built-in SynthID verification verdict in the generated CSV.

The harness shares one latent-noise field across the sequence to avoid adding independent frame noise. Its temporal-residual metric is a fidelity check, not a watermark detector.

The 2026-07-29 two-carrier calibration produced genuine built-in verifier results: both matched controls were positive, the stronger candidate was negative on both carriers, and a weaker candidate was negative on one. On 2026-07-30, adversarial follow-up prompts returned UNAVAILABLE after asking ordinary Gemini to ignore and reinterpret the detector result. Those follow-ups were mistakenly treated as a stricter oracle; they were not detector reruns. The implementation remains exposed as video invisible and remove_video_invisible. Its oracle-certified default is the product operating point. A fresh control-positive, candidate-negative pair remains an optional per-file audit after provider changes or for unusually important files.

The 2026-07-31 full-clip check added a public eight-second Veo carrier. The source and the complete 0.10 product output were both detected, proving the surrounding resize / frame-rate / codec path had not silenced the oracle. The complete 0.15 product output was not detected, so 0.15 became the default. The tracked source and output hashes, fidelity metrics, and verdicts are in data/evaluations/video-synthid-oracle.csv; generated media remains untracked.

Tier E -- robustness and adversarial inputs

Malformed and hostile inputs, including truncated files: at many offsets, corrupt headers, 16-bit and CMYK, absurd dimensions, decompression bombs, zero-byte files, unicode and RTL filenames, symlinks, read-only output dirs, concurrent runs on one file. The bar is never "handles it" but never raises and never silently degrades.

Build order

1. A1 sidecar regression -- highest value per hour, unattended, needs no new labels.
2. A2/A3/A4 parity and invariants -- representative local set, reusing existing audit scripts.
3. B1 fill quality -- closes the oldest unmeasured claim in the project.
4. B2 detector curves -- cheap, and directly guards the geometry class of bug.
5. A5 contract sweep over a representative local set.
6. B4 resource ceilings, E robustness.
7. C recall expansion -- gated by labeling appetite.
8. D oracles -- manual, per release.

Every tier writes a versioned snapshot so runs are comparable over time; a run that cannot be diffed against the last one is a one-off, not a regression suite.

What the measurements imply for detection work

Recorded here because each item is grounded in a number from this campaign, not because it is a prioritized plan (that lives elsewhere -- see the note at the end of this section).

Metadata absence does not disable the detectors -- it disables the RELAXATION

The detectors are pixel-based and need no metadata. What metadata does is relax the false-positive gate (auto vs strict). So "work better without metadata" means strengthening the strict-path detectors themselves; it is not a gating problem.

Per mark, what actually goes away when metadata is stripped:

• The pill loses its metadata-confirmed arm. Without metadata it depends on the bottom-right Jimeng wordmark.
• The sparkle already runs on pixels. Its threshold remains a deliberate recall-versus-precision trade.
• Uncovered vendors are metadata-independent gaps. They require a detector of their own.

Where the evidence points

1. Prefer per-vendor CJK templates when fitted geometry separates similar marks.
2. Recalibrate every detector when changing its front-end. Thresholds do not transfer between binary, tophat, and gray.
3. Treat the Jimeng wordmark as a load-bearing confirmation path for the pill.

Measure before improving

Use Tier B2 detector-response curves before tuning a detector with sparse labeled examples. Sweep size, contrast, aspect, and background texture so geometry regressions are visible without exposing private evaluation statistics.

This section records what the measurements imply technically. Prioritization is tracked separately, outside this repo.

Local end-to-end verification

Run scripts/real_examples_e2e.py against representative local inputs before releases that affect image handling. The script must read from .local-eval/, write only untracked temporary output, and report behavior without exposing dataset provenance or aggregate private measurements.

The cheap video seam is covered automatically by TestVideoVisibleFullClip::test_removes_complete_clip_and_preserves_sequence_and_audio. It constructs a full synthetic Sora-like MP4 with AAC audio and C2PA provenance, then exercises detection, temporal arbitration, OpenCV fill, real ffmpeg encoding, metadata stripping, audio stream copy, and atomic publication through the public API. A separately encoded clean control supplies the paired frame-to-frame deltas inside the filled region, so the gate detects temporal flicker rather than treating all output motion as an error. The default motion-compensated fill must score strictly below a separately encoded frame-local opt-out on the median paired error, while its high-percentile error cannot regress. The dedicated Linux CI job installs ffmpeg explicitly. Keep real-provider and learned-backend sequence evaluation local because those inputs or model downloads do not belong in the core matrix.

The local real-provider audit runs one complete clip for every registered video mark. OpenCV and MI-GAN must remove every accepted frame, leave the second-pass detector quiet, preserve source stream starts and duration, and copy AAC packets exactly when present. Run one complete LaMa clip to verify wiring and resource tier; its CPU throughput makes a six-provider online matrix counterproductive. Store generated outputs and the detailed CSV only under .local-eval/.

The audio side of that contract now has a local matched-oracle study: Audio provenance experiment embeds AudioSeal into synthetic carriers, verifies the copied-audio bitstream identity and a watermark verdict invariant across remove_video_visible, and measures the audio-path attacks separately. Its carriers are synthetic non-speech audio and its oracle is AudioSeal, so it validates the copy semantics, not provider audio marks. The audio and video arms are now first-class benchmark rows: benchmark kernel accepts media_type: audio and media_type: video with the revision-pinned audioseal and videoseal adapters, and both cohorts carry matched negatives, codec, and degradation arms as strict manifest cases. The temporal layer is closed too: videoseal detection records carry per-frame bit accuracy, and VideoSeal temporal evaluation measures the aggregation matrix and the crf decay curve on synthetic and real clips. What remains open here is a wider real-content sweep and the learned temporal inpainting comparison.

The privacy-tiered C2PA resolution item is closed as a product decision: the resolution research measured the five vendors' resolver contracts (payload-only, asset digest, or unpublished), grounded them in C2PA 2.2 section 18.10.5, and concluded that this tool resolves nothing over the network at any tier - querying a resolver tells the vendor someone is inspecting the specific content, which is a leak in the opposite direction of a removal tool's purpose. A registry guard test pins the published-API set so ecosystem changes surface consciously.

The external-benchmark item is closed as a licensing verdict: the investigation verified ETI and W-Bench against their primary sources - ETI carries no license anywhere and stays off limits without organizer permission, while the W-Bench dataset is MIT with the VINE code non-commercial. A W-Bench profile is therefore buildable on our kernel; whether it is worth building is a scoping decision that now has no legal dependency.

The same full-clip gate runs the public metadata-only path against its real MP4 and verifies unchanged file size, decoded frames, stream properties, and AAC packets. A separate synthetic large-mdat test rejects full-source read_bytes(), hashes the copied media payload, and mutation-checks both C2PA and TC260 survival.

A companion full-clip VFR case alternates three frame durations, runs the public visible-removal API, and compares every output display timestamp to the source within one source time-base tick. Its source starts at a non-zero PTS, so the test also verifies retained video/audio stream offsets, container duration, and copied AAC identity. Mutations that disable the timestamped NUT bridge or reset its start PTS must fail this gate. A separate constant-rate clip with the same non-zero start guards the start-offset routing without relying on the VFR branch.

Signal discovery cadence

The project skill signal-discovery orchestrates the corpus audits in this plan. scripts/corpus_gap_scan.py is only the metadata-serialization probe inside that workflow. It now records raw markers, input shapes, identify errors, integrity clashes, unattributed results, and blind-marker candidates, but it does not cover visible marks, proprietary invisible marks, detection/removal parity, or historical crossfire.

Run a fresh-delta discovery review weekly after the retained corpus is refreshed, using corpus_gap_scan.py --since with an overlap from the preceding run. Replay the complete historical corpus before releases that change a detection, metadata, container, registry, invisible-routing, or removal seam, after every detector registration or recalibration, and monthly as a backstop. The historical pass is required because newer code can expose an old serialization and a new detector can crossfire on an old neighbouring cohort. A stratified sidecar sample is a useful preflight but does not replace a triggered full replay.

Removal-parity sweeps are change-triggered rather than weekly: run the matching metadata, visible, video, or invisible audit whenever either its detector or cleaner moves. External provider oracles are claim-triggered, never calendar-triggered. Keep private inputs and reports untracked, and implement an actionable detector or cleaner gap in this library rather than in a downstream consumer.

Standing gap

None of this is in maintain.sh, and it should not all be -- the sweeps take hours. But that means no detector-accuracy or CLI-contract regression is caught automatically today. The endpoint of this plan is a cheap subset (fixtures-only smoke + a sidecar diff on a fixed 500-image slice) that CI can run, with the full sweeps staying pre-release.