The resolution curve's 0.154 top left the 4.33 MP CJK-sign fixture
SynthID-detected x3 in Gemini on the full production path (visible
stage -> qwen-zimage seed 0 -> resize-back -> metadata strip), with a
valid pixel-identical stripped control in the same session
(2026-08-18). Google-provenance content now resolves to the flat
QWEN_ZIMAGE_GOOGLE_STRENGTH 0.30 floor instead of the area curve -
anchors at 0.30 measured clean in Gemini on two fixtures (CJK sign +
18-face) at 3/3 checks across two work accounts, and stayed clean
under the vae-glyphs donor layer. openai/unknown content keeps the
curve; an explicit strength still wins.
The engine forwards the new kwarg into WatermarkRemover.remove_watermark,
which rejected it with TypeError on the real (non-fake) path - caught by
the deployed Modal smoke test, not by the unit fakes.
The whole-frame 15% Qwen-VAE blend returned detector-visible OpenAI
SynthID on poster-scale manifests through the engine text-manifest
path (official Content Provenance API, 2026-08-19: restored detected
x6 with the anchor, clean x6 without it; base outputs clean x6;
pixel-identical stripped controls detected, proving the pixel channel).
Add fidelity_anchor=False to remove_watermark and InvisibleOptions and
--fidelity-anchor on the CLI to reproduce the 0.27.0 research
behavior. Text-box MAE cost of the new default is under one point on
all three fixtures (11.60->11.72, 7.79->7.86, 7.57->8.13).
Issue #84: text, tables, and UI screenshots are the worst case for
--force regeneration. Document that the default pipeline skips the
invisible stage without a detectable signal, that visible/metadata
stages never redraw glyphs, and why paste-back is not offered.
The finding lived only in the research archive, whose own banner says it may
describe prototypes that were later removed. It belongs where users read
limitations: the path strips every metadata marker while copying the audio
bitstream byte for byte, so a clean local report on a clip with generated audio
is unproven rather than a guarantee.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Everything here is evidence and state, not behavior: the shipped operating point
is untouched at 512 px / 12 fps / noise_std=0.15.
The carrier was downloaded and its sha256 matched the manifest, so its properties
are now measured rather than assumed: 1920x1080 at 24 fps with an AAC audio track.
Those three fields go into the two 2026-07-31 rows, which could not previously
tell a reader what the downscale factor even was.
The geometry prize is measured end to end for the first time. Against the
untouched source, 1024 scores +3.46 dB over the shipped 512, while the entire
noise_std axis is worth 1.92 dB. Two readings that the table alone hides are
recorded with it: the temporal residual IMPROVES with resolution, because the
shared noise field lives on the latent grid and is four times finer relative to
the frame at 1920; and the frame-rate arm cannot be judged by these metrics at
all, since they price its cost and not the smoothness it buys.
Two findings that outrank the quality question. The certified row does not
reproduce -- a rerun of the same configuration on a different device and dtype
gives a different hash and different metrics, and the manifest records neither.
And the pipeline copies audio byte for byte while stripping every metadata
marker, so it can emit a file our own detector calls clean with an untouched
Google-generated audio track inside it. The mechanism is proven on two carriers;
whether that audio carries a mark the verifier reads is not, and only the oracle
can say.
The six prepared oracle submissions were never run: file upload to the verifier
failed at the tool level. Their artifacts are gone with the scratch directory,
which is the intended lifecycle for generated media, and the document says what
rebuilding costs.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
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>
Renaming it strands every pip install and every pinned dependency on it,
including the ComfyUI node package, and the console script has the same
property since users have it on PATH. Recorded at the name itself, where
whoever considers changing it will be looking.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
These directories are empty today, so git shows nothing and the gap is
invisible. The moment the sync writes into one, a routine running
git add -A commits another harness's generated config, and the harness
audit then fails after the fact rather than before it. Paths are the
canonical untrackedConsumerPaths list, anchored to the repo root so a
same-named source file deeper in the tree is not caught.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
All four plugins it enabled are enabled globally, and its allow list says
nothing under the global defaultMode: bypassPermissions.
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>
Updates the three version sources the release doc names -- `pyproject.toml`,
`__init__.py`, and the root package entry in `uv.lock` -- and carries the marker
simplifications uv produced when it re-resolved the lock.
The release itself is not started here: the tag, push, and GitHub Release are the
remaining steps, and PyPI publishing triggers on the published Release rather than
on a tag push.
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>
`InvisibleOptions` promises in its docstring that every default mirrors
`InvisibleEngine`. Two fields made that promise cost something to keep: `force` is
not an engine parameter at all, and `controlnet_scale` was a third spelling of the
engine's `controlnet_conditioning_scale`. The mirror test carried an exception
table for each. This removes both, so the comparison needs no exceptions -- a field
that needs one is a field that belongs somewhere else.
`force` decides WHETHER the engine runs, which is settled before it is built, so it
joins `backend` and `sensitivity` as a parameter of `remove_all` and `remove_batch`
and is threaded to `_run_invisible` as its own argument. `controlnet_scale` takes
the engine's own name; the click option stays `--controlnet-scale` and is now
translated exactly once instead of at three forwarding sites.
Safe to do today: both symbols landed after 0.25.0 and have never been published.
The forwarding turned out to be the weaker half. A defaults comparison cannot see a
hardcoded literal at the seam, and `_run_invisible` passed the entire suite with
`controlnet_conditioning_scale` pinned to a constant. Each of the two knobs also
reaches the engine through TWO paths -- `remove_all` versus `remove_batch(mode="all")`
for `force`, `_run_invisible` versus `_batch_engine` for the scale -- and guarding one
left the other free to hardcode with a green suite. So:
* `test_every_field_arrives_at_the_engine_with_the_caller_s_value` drives the real
seam with all 13 fields set off their defaults; mutating any one of them to its
default fails it.
* `test_force_reaches_the_scrub_gate_in_every_scrubbing_mode` and
`test_batch_controlnet_scale_flows_to_the_cached_engine` are parametrized over
both modes, so neither path can be pinned alone.
Also fixes an order-dependent test surfaced by the added tests reshuffling the xdist
shards. `test_visible_path_decodes_file_once` counted every `image_io.imread` in the
process, but the Gemini engine loads its own bundled capture assets on first
construction, so the count was 3 on a cold engine and 1 on a warm one and the test
passed only when an earlier test happened to build the engine first. It now counts
decodes of the SOURCE, which is the invariant it exists for, and still fails when the
shared decode is broken. The production path was never wrong: the source bitmap is
decoded exactly once.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Every code-referencing claim in the docs, the README and the rules files was
checked against src/, and each finding was re-derived independently before it
was applied. 35 held, 5 were false positives.
Two of them were code, not text. `InvisibleOptions` promises in its docstring to
mirror `InvisibleEngine`, and two defaults had silently stopped:
`max_resolution=None` reached `_target_size`'s `max_resolution > 0` and raised
`TypeError` on every library call that left the options alone, and
`cpu_offload=True` made a library run slower than the identical CLI run. Both are
fixed, and `TestInvisibleOptionsMirrorTheEngine` compares the two signatures
field by field rather than pinning the two values that happen to be known. A
companion assertion in `TestTargetSize` reads the engine's own declared default,
so a drift on the engine side -- which the mirror check alone would accept,
because both sides would still agree -- fails too.
The user-facing docs: README called `invisible` GPU-optional where it raises
without CUDA, and gave the image `metadata` command `video metadata`'s output
rule, promising the source survives a command that overwrites it. Yuanbao was
missing from the supported-mark list. `veo` was listed among the video policies
that require a run anchor, though its row sets no `anchor_iou`.
`known-limitations` called ControlNet the default profile and contradicted
itself ninety lines below. An unescaped pipe truncated the `hailuo` table row.
The `dev` extra, the CI shape, ffmpeg's role, the sdist boundary and the
strength-curve range were corrected, and `remove_all`/`remove_batch`, the pill
gate, `erase --keep-metadata` and `all`'s CUDA failure mode were documented.
Research notes that described removed modules, extras and flags in the present
tense now say so once in the page banner instead of sentence by sentence, which
covers the whole page rather than the lines that happened to be noticed, and one
fixture is referred to by role rather than by name.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The visible-mark path had grown three copies of one ladder sweep, four
near-identical `detect` arms, and four hand-rolled `footprint_mask` overrides;
mark knowledge sat in five hand-maintained tables across three modules; and the
flagship `all`/`batch` pipeline existed only in cli.py, written twice with
divergent behavior.
Detection is now one measurement. `_ladder_best` replaces the three sweeps,
`_scan`/`_verdict` replace the four arms, and the winning box travels to the
mask on `TextMarkDetection.match_box` instead of being swept a second time.
`detect_both` returns the strict and relaxed verdicts from one scan, which
halves the arbiter's perception cost (260 -> 130 matchTemplate calls on a 2048²
image, verdicts identical field for field). A per-mark demotion goes in the new
`_post_gate` hook, never in a `detect` override -- an override is invisible to
the single-pass path, which is how the RunningHub and Yuanbao anchor gates
briefly stopped applying.
Everything about a mark is now one registry row: product, label regime, the
platform sentence `identify` reports, the metadata signals that confirm it, and
its TC260 producer codes. `identify._VISIBLE_MARK_PLATFORM`, the signal mapping
in `api.visible_provenance`, `_PRODUCT_OF` and the pill veto are derived from
those rows.
`api.remove_all` / `api.remove_batch` are the library form of the `all` and
`batch` commands; the CLI is a wrapper that owns console text and exit codes.
Progress is a `(stage, detail)` pair of stable tokens, so the CLI keys its
wording off structure rather than parsing the library's prose back.
Two intentional behavior changes, both verified against a recorded 811-image
sample of detector verdicts, removal-mask hashes, arbiter decisions and
`identify` reports:
* A TC260 label now relaxes the vendor its `ContentProducer` names rather than
ByteDance's pair on every China-AIGC image. 333 of 811 samples move; on 185
of them the previously relaxed pair was simply the wrong vendor, and the
mark actually present never reached the relaxed gate its own
`provenance_ncc_factor` was calibrated for.
* A confident LibLibAI detection suppresses the Jimeng pill, like every other
TC260 product's mark. It was registered alongside RunningHub and Baidu, both
of which were added to the hand-written veto list, and it was not. 1 sample
moves, and it is exactly the co-firing case.
Nothing else in that record changes: detector verdicts, mask hashes and
`identify` verdicts are byte-identical, and all 200 calibration constants are
untouched.
Also: `aigc_label` and friends plus `extract_c2pa_info` are memoized on
(path, mtime_ns, size) -- size because this package rewrites in place; the
native TC260 container readers route on magic bytes instead of the file
extension, so a mislabeled AVI or FLV is no longer invisible; `identify` shares
one pixel decode between the DWT-DCT and visible stages (TrustMark keeps its own
Pillow decode, which is not substitutable); and the six `stabilize_*` video
wrappers collapse into one policy table.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Patch and minor only: cffi, coverage, hf-xet, modelscope-hub, typer. The majors
the resolver held back (numpy 2, opencv 5, tokenizers 0.23) are the ones
.github/dependabot.yml already documents as blocked upstream.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
