Review triage

Type: kb/types/instruction.md

Inspect note-changed stale pairs and acknowledge those where the diff does not invalidate the existing evidence for that criterion. This reduces the assay queue before launching workers. For closed-ended verdict pairs, ack carries the outcome forward; for open-ended report pairs, it only reuses the report as fresh evidence and does not imply that its findings were handled.

Inputs:

  • {model-partition} — the freshness partition whose existing evidence may be carried forward
  • {criteria} — gate ids, bundle names, conformance requests, or the opt-in critique report assay; --all-gates selects verdict criteria only
  • {note-scope} — one or more note paths or directories to limit scope

Steps

1. Capture note-changed candidates

commonplace-review-target-selector --model-partition {model-partition} {criteria} --note {note-scope} --json --reason note-changed > {ack-manifest}

If the output object has "targets": [], stop — nothing to triage. Keep this exact file: its baseline revisions and current hashes bind the later acknowledgement to what you inspect.

2. For each note-changed pair, judge the diff

For each entry in targets, read reasons, every item in changed_inputs, criterion_id, and result_kind. A joint note-and-criterion edit is returned by the note filter but still carries both changes. Ask whether each diff invalidates what the criterion's existing result says.

Guidelines:

  • A typo fix, whitespace change, or link-text tweak is insignificant for almost every gate.
  • A rewording of a claim is significant for semantic/grounding-alignment and semantic/internal-consistency but probably not for structural/general-before-specific.
  • Adding or removing a section is significant for prose/proportion-mismatch and semantic/completeness-boundary-cases.
  • Adding or removing a source citation is significant for prose/orphan-references and semantic/grounding-alignment.

When in doubt, do not ack — rerun the assay.

3. Retain only authorized observations

Edit {ack-manifest} so targets contains only pairs authorized for acknowledgement. For a joint edit, retain every harmless changed_inputs item and its corresponding reason; remove a role that must stay stale for rereview. Do not update hashes, snapshot ids, paths, or baseline revisions.

4. Ack the inspected candidates

commonplace-ack-review --input {ack-manifest}

This advances only the retained input roles while preserving the evidence review pair. An intervening file edit or baseline transition fails instead of accepting uninspected state. It produces no new judgment and does not rely on touch or filesystem timestamps.

5. Report

Report which pairs were acked and which were left for review. The remaining stale pairs will be picked up by the next review sweep.