Q6 is a reproducible benchmark for photosensitive-epilepsy (PSE) detection in video, plus a detector implemented from the text of the standards (WCAG 2.2 SC 2.3.1, Trace24, ITU-R BT.1702, Ofcom GN2 Annex 1, NAB-J, ISO 9241-391) — not tuned against benchmark labels.

It exists because, despite PSE being a well-studied harm (the 1997 Pokémon Porygon incident triggered seizures in ~700 viewers in Japan and is what spurred most of the regional broadcasting rules), there's no public, cross-standards, cross-tool benchmark. We built one.

Q6's classical detector evaluated on TRACE's public corpus (pse-test-media, BSD-3) — same fixtures, same labels, every tool running through an identical harness with strict adapter / label isolation:

† marks tools whose underlying interpretation of the WCAG area clause is materially looser than the WCAG-strict reading TRACE labels assume. IRIS implements Harding-classic (~87,296 px floor on contiguous active area); Kaya 2025 reads "any 341×256 rectangle" as "any 1/3 × 1/3 frame cell" (~57,600 px floor on 1920×1080). WCAG-strict puts the floor at ~21,824 px via a sliding reference rectangle. A low MCC under those rows reads as "the tool targets a looser interpretation than the labels assume" rather than "the tool is broken." Full diagnostic in detector/ml/SANITY_CHECKS.md Checks 3 and 5.

A learned detector stacked on Q6's per-frame trace (A6 in our sweep — 5-fold CV on TRACE) reaches MCC +0.355, a +0.155 absolute lift over the classical baseline on the same 293-sample subset, and the only non-zero MCC any published ML PSE detector reaches on this corpus.

Methodology and per-approach numbers live in detector/ml/EXPERIMENTS.md. Cross-checks on the ML headline and one real adapter bug we caught in vf_photosensitivity are in detector/ml/SANITY_CHECKS.md.

A multi-resolution structure capturing where, when, and why content is hazardous, plus how to fix it:

• Verdict + continuous severity score at fixture level

• Per-axis aggregates (luminance / red / pattern / count): verdict, peak windowed transition count, peak hazard area, fail intervals, margin to threshold

• Per-frame hazard regions — a list per frame, each with bbox, centroid, set of triggered hazard classes, per-class severity, confidence band (marginal / clear / severe), structured mitigation hints with alternatives, linked standards clauses, and a counterfactual ("would PASS if...")

• Per-class swimlane chart — one-glance answer to "did this video have hazards, of which class, and when":

  

  (60 fps, full-screen 31 Hz flash. The luminance lane carries a red FAIL chip and twelve interval bars at the burst times; the red-flash lane shows a green PASS chip and stays empty. Lanes are PSE-safe by construction: bars are continuous spans that cannot flicker, severity is encoded inside each bar via fill darkness, and the verdict chip in each lane header makes a single-frame FAIL just as visually loud as a sustained one.)

A per-fixture HTML report wraps all of the above into one self-contained file suitable for content creators / accessibility reviewers.

git clone https://github.com/qwertey6/Q6
cd Q6
pip install -r requirements.txt

# self-test the detector (no corpus needed)
python3 -m detector --self-test

# render a per-fixture HTML report on a sample video
python3 -m report.fixture_report path/to/video.mp4
# → report/out/video.html + video.heatmap.png

# full reproducible benchmark (downloads corpus, runs all adapters, scores)
make all

make all produces results/scores/scores.{json,csv} and per-fixture HTML reports under report/out/. make corpus / make harness / make report run the milestones independently.

Each profile implements the standard's text directly; thresholds are cited clause-by-clause in detector/THRESHOLDS.md:

• WCAG 2.2 SC 2.3.1 (W3C accessibility) — strict reading
• WCAG 2.2 classic — Harding-rectangle area reading (looser)
• Trace24 — TRACE's proposed stricter guideline (Jordan & Vanderheiden, 2024)
• ITU-R BT.1702 — international TV broadcast baseline
• Ofcom GN2 Annex 1 — UK broadcast
• NAB-J / J-BA — Japan (post-Pokémon Shock guidelines)
• ISO 9241-391 — ergonomics / accessibility cross-cut

Per-fixture labels for the first five come from TRACE's pse-test-media corpus. NAB-J labels are derived analytically (see corpus/derive_per_standard_labels.py) — this is internal consistency, not external validation; the script documents the derivation honestly.

corpus/        labeled fixtures (TRACE/pse-test-media + 45 Q6-extended synthetic)
                + MANIFEST.csv with per-standard label columns
                + PROVENANCE.md (how the corpus was constructed)
detector/      Q6 detector (classical + ML)
                + THRESHOLDS.md (every numeric constant cited to a clause)
                + ml/ — A1–A6 experiment sweep + writeups
harness/       benchmark runner, adapters (Q6, IRIS, ffmpeg, flickerfilter,
                Apple VFR, Q6 MLP), normalized-result schema, scoring
report/        comparative report + per-fixture HTML + heatmaps

Adapter / label isolation is the non-gameable invariant: the runner reads only fixture paths from the manifest, never labels; scoring joins to labels in a separate pass. No adapter has privileged access.

PolyForm Noncommercial 1.0.0. Personal / academic / public-research / public-health / government use is granted. Commercial licensing is available by contact.

Third-party components retain their upstream licenses; see THIRD_PARTY_NOTICES.md.

• detector/THRESHOLDS.md — every numeric threshold + its clause citation; open-questions log
• detector/ml/EXPERIMENTS.md — A1–A6 ML sweep, methodology + per-approach numbers + conclusions
• detector/ml/SANITY_CHECKS.md — cross-checks on the ML headline; real adapter bug found and fixed
• corpus/PROVENANCE.md — how the corpus was assembled, what's where, how to reproduce
• PLAN.md — original architecture document (mostly historical now)

External contributions are welcome — please open an issue or draft PR before substantial changes. A Contributor License Agreement (CLA) is required for non-trivial contributions; light fixes don't need one.

This repository is engineering tooling, not legal advice. Q6's verdicts should not be the sole basis for production-content compliance decisions. Before any public publication of benchmark results or commercial release, the licensing and any redistribution of generated/derived media should be reviewed by counsel. We claim no endorsement from any upstream project or standards body.