triage-dailieslisted

# triage-dailies Drive the dailies morning ritual: report what the night produced and what it cost, look at the evidence behind every doubtful verdict, get the human's ruling, bank it as gold, keep the judge calibrated and honest, and end with a concrete plan for tonight. ## Ground rules - You recommend; the human rules. **Never run `dailies gold add` without an explicit human ruling on that specific take. Never derive a label from the judge's verdict, from your own recommendation, or from the user's silence. The `gold` block belongs to the human (SPEC.md); review tooling and agents must never write or infer it.** - The JSON shapes below are hard-coded from `cli.py` and pinned by `tests/test_skill_envelope.py`. Trust them over any prose. - All state lives in `<clip>.take.json` sidecars next to the clips (shape in SPEC.md). Write through the CLI, never by editing sidecars; sidecar tools must preserve keys they do not own. ## 1. Judge health first Before trusting any verdict, read `dailies-judge-check.json` (in the takes directory, else the working directory). Shape: {"runs": [..., {"engine": <str>, "n": <int>, "agreement": <float>, "kappa": <float>, "false_kills": <int>, "missed_kills": <int>, "created": <ts>}]} The last run is the judge's report card. Flag it when: the file is missing, kappa is below 0.6 (tonight's `watch --regen` will refuse to start), or the newest run predates a judge model, rubric, or generator change the