motion-evidence-review
Skill78 starsChanged 9 days agoApache-2.0View on GitHub
How to read beep-qa motion evidence: correlate frame strips, contact sheets, and GIFs against events.ndjson timelines, map frame timestamps to action events, and cite frame+event pairs when judging or fixing. Use when reviewing QA evidence, writing a vision-judge finding, reproducing a finding from a session's artifacts, or auditing an old round's evidence in a goal packet or PR.
In this repository38 other files it gives its agents.
- AGENTS.mdAGENTS.md
- explorations/AGENTS.mdAGENTS.md
- adhdSkill
- atom-reactivity-specialistSkill
- browser-qa-loopSkill
- closeoutSkill
- codex-findingsSkill
- crispenSkill
- effect-first-developmentSkill
- exploreSkill
- grillingSkill
- grill-meSkill
- grill-with-docsSkill
- jsdoc-annotation-specialistSkill
- onepassword-secret-refsSkill
- ontology-architectSkill
- ontology-conceptualizerSkill
- ontology-curatorSkill
- ontology-foundational-auditorSkill
- ontology-mapperSkill
- ontology-requirementsSkill
- ontology-scoutSkill
- ontology-validatorSkill
- operator-handoffSkill
- orchestrateSkill
- plasma-sensitive-entrySkill
- portlessSkill
- qa-session-opsSkill
- quality-review-fix-loopSkill
- reflectSkill
- remote-development-agentsSkill
- schema-first-developmentSkill
- shadcnSkill
- sparql-expertSkill
- teachSkill
- turborepoSkill
- unslopSkill
- yeetSkill
The file
Files
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)
- Event clock — witness
tEpochMs=performance.timeOrigin + performance.now(), wall-clock epoch ms. Every line ofevents.ndjson. - Video time — seconds from the recording's first frame.
- The mapping —
session.json → clockSync:videoT = slope·tEpochMs + offsetMswithmethod(beacon|obs-record-state|assumed-start) andconfidence. For a beacon fit,highrequires 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 whileevents.ndjsonseq 141→158 shows an active pointer drag (pointer-downseq 141 at tEpochMs …, nopointer-upuntil seq 158) — selection-smear during sash resize.
Rules:
- Cite artifact path + frame index/range + event
seqrefs. InventoryQaEvidenceRefhas 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
session.json→ scenario + commit (commitSha,commitDirty).- The finding's
eventIds→ grep those seqs inevents.ndjson→ exact gesture path, coordinates, timing. - Re-run just that scenario (
.beep/qa-capture.mjsscenario name) or replay the gesture manually in Lane B at the recorded coordinates. - 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.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

