Repair image and metadata publication contracts, detection routing, research
artifact validation, release verification, and the published agent skill.
Refresh the dependency lock and reconcile documentation with verified behavior.
Add regression coverage and record the completed audit and corrected VideoSeal
measurements. Preserve the existing Accelerate and Lightning versions.
Validation: Ruff and Pyright pass; 2358 tests pass with 3 skipped. The maintenance
gate exits 2 on the documented Accelerate and Lightning security findings.
The Content Seal manifest named the production GPU worker in its provenance
column and in two row keys, several docstrings described their sources as user
uploads from the private corpus, and that corpus was called by the storage
product holding it. Each is now what a public reader can check: a pipeline and
strength, material that is not publication-cleared, and a private corpus. The
corpus test follows the renamed rows. Deployment and tier detail goes the same
way, since which backend each hosting tier can afford is a product decision,
not a library fact.
Twelve harnesses that need a cloud GPU account leave scripts/. The research
pages that cite their findings name them as harnesses kept outside this
repository and stop implying a reader can rerun them; the tracked inputs and
recorded verdicts remain the durable record, and the offline analysis step
stays here with its test. A development invariant states the rule so the next
harness is not added back.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The packaging tests a catalog runs say nothing about content, so the skill's
prose drifted from the CLI exactly the way an unguarded seam does. These tests
read the live Click tree: mark keys, flag choices, batch modes, pipeline
profiles, vendor cohorts, exit codes and extras must match what the skill says.
Guard the SET, not the instances. Per-flag comparisons only check what the skill
already documents, so `--vendor`, `--pipeline auto`, and then the whole
`classify` and `verify-openai-synthid` commands each shipped with a green suite.
`test_every_cli_command_is_named_somewhere_in_the_skill` closes that.
None of it proves the only thing a skill is for: that a model reaches for it and
then obeys it. `scripts/skill_eval.py` builds a throwaway project with the skill
installed and a recording stand-in for the CLI on PATH, runs a headless agent on
one user prompt, and grades the recorded command trace. Grading is mechanical --
"did it run `all` where the probe said nothing is written" is a fact in the
trace, and "did it name an output file that is not on disk" catches a fabricated
result without asking a model to judge itself. Agents are nondeterministic, so
the report is a pass rate per check over repeats, never one verdict.
The harness is tested too: the stand-in's exit codes are compared against the
real CLI where the state is reproducible locally, the grader runs over hand-built
traces, and every case is checked for internal consistency, because a scenario
whose `identify` names a mark its `visible` cannot find measures the case rather
than the skill.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
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.
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>
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