audit

/audit

Intent. Check a note's quality against the contract. Three orthogonal checks: schema (frontmatter), links (no danglers), cold-read (can the model predict the body from title+description+frontmatter alone?).

Modes

Mode	Schema	Links	Cold-read	Topology
--mode=quick	yes	yes	no	no
--mode=full (default for single note)	yes	yes	yes	no
--vault	yes	yes	sample	yes

Behaviour

Single-note audit (default)

1. Schema check. Run tools/schema-check.py <note>. Validates frontmatter against the active preset's contract. Reports missing required fields, invalid enum values, descriptions that are too short / too long / merely paraphrase the title.
2. Link check. Run tools/link-check.sh <note>. Reports [[wikilinks]] whose target file doesn't exist (danglers) plus a list of context-prose-less inline links (bare [[a]]. [[b]]. with no surrounding sentence).
3. Cold-read. Read only the title + frontmatter + description. Predict what the body should say (3 bullets). Then read the body. Compare. Report:
   • Strong: body matches prediction in spirit
   • Weak: body adds material the description didn't trail-blaze (description should layer heuristic → mechanism → implication; if the body has a key idea the description omits, the description is too short)
   • Wrong: body contradicts the description
4. Synthesis. Print a short report:

   /audit notes/claims/X.md
     schema: 2 issues
       - description is paraphrase of title (not adding info)
       - confidence missing
     links: 1 dangler
       - [[some-claim-that-doesnt-exist]]
     cold-read: weak
       - body introduces "scaling regime" not flagged in description
     fix priority:
       1. rewrite description to trail mechanism (high)
       2. set confidence (medium)
       3. resolve or remove [[some-claim-that-doesnt-exist]] (medium)
   

Vault audit (--vault)

1. Same schema + link checks across every notes/**.md.
2. Run ../walk/tools/topology.py orphans and topology.py moc-coverage for structural health.
3. Sample cold-read on ~10 claims (random + recent) — full cold-read on every note is too expensive.
4. Run tools/health-metrics.py to print the four-family health dashboard (coherence, connectivity, boundary precision, confidence distribution). This is the per-session eval signal from the claim-fact-gradient framing — print it as the ## Health Metrics section of the report.
5. Bridge-map consistency check (_cross-domain-bridges.md):
   • Every synthesis claim listed under ## Wired bridges must exist on disk and have type: synthesis.
   • Every synthesis claim that bridges a pair surfaced by topology.py disconnected-clusters (with both clusters ≥ 5 members) should appear in the bridge-map. Flag synthesis claims that meet the bridge criterion but aren't listed — they're un-indexed bridges.
   • Every meta-MOC (frontmatter meta: true) must have parent_map: "[[index]]" and a scope field. Other meta-MOCs are rare; flag any beyond _cross-domain-bridges for human review of whether the meta-level is justified.
6. Synthesise: total counts, top-N worst-offenders, suggested order of fix.