motion-evidence-review

Motion Evidence Review

Three roles consume QA evidence at different times: the vision judge (via its prompt), the fixer reproducing a finding, and PR/packet reviewers auditing after the fact. This skill is the shared literacy. The loop protocol and the inventory/severity format live in browser-qa-loop — this is how to READ.

The three clocks (get this right first)

1. Event clock — witness tEpochMs = performance.timeOrigin + performance.now(), wall-clock epoch ms. Every line of events.ndjson.
2. Video time — seconds from the recording's first frame.
3. The mapping — session.json → clockSync: videoT = slope·tEpochMs + offsetMs with method (beacon | obs-record-state | assumed-start) and confidence. For a beacon fit, high requires residual RMS ≤25 ms and at least four matched pairs. RMS is not a maximum per-event error or a frame-accuracy guarantee. Inspect the method, timestamps and extraction padding before making timing claims; see the method-specific table below.

references/frame-event-correlation.md walks the arithmetic with worked examples; references/evidence-anatomy.md decodes file naming, strip ordering, and sheet grids.

Reading a finding (or writing one)

A well-formed claim always pairs pixels with events:

Frames 12–15 of frames/drag-sash--strip/ (t≈3.31–3.41 s) show a native selection highlight across the neighbor panel while events.ndjson seq 141→158 shows an active pointer drag (pointer-down seq 141 at tEpochMs …, no pointer-up until seq 158) — selection-smear during sash resize.

Rules:

• Cite artifact path + frame index/range + event seq refs. Inventory QaEvidenceRef has fields for exactly these — fill them.
• A defect in ONE mid-gesture frame is real even when the final screenshot is clean; the converse also holds — a single odd frame at LOW clock confidence near a window edge may be a neighboring gesture. Check the event seq range before claiming.
• What a GIF proves: a human-visible summary. What it cannot prove: exact timing (GIF fps ≠ source fps, frames are sampled). Timing claims come from strips + timeline, never from GIF playback feel.
• Markers (kind: "marker") are semantic labels — scenario boundaries and agent annotations. Use them to scope which events belong to which gesture.

Reproducing a finding

1. session.json → scenario + commit (commitSha, commitDirty).
2. The finding's eventIds → grep those seqs in events.ndjson → exact gesture path, coordinates, timing.
3. Re-run just that scenario (.beep/qa-capture.mjs scenario name) or replay the gesture manually in Lane B at the recorded coordinates.
4. Any extracted frame's own XMP-beepQA provenance tags (inspect with exiftool) identify its session if files got separated from the round dir.

Auditing old evidence

inventory.json is schema-validated (beep qa judge-lint --round N re-checks any round). judge/manifest.json lists what the judge actually saw INCLUDING dropped — an inventory is only as complete as that manifest; treat dropped evidence as unreviewed, not clean.