Deep audit

The owner typed /vault-deep-audit, which includes /vault-audit: both are authorised for this run (AGENTS.md § Compile and Audit, D78). Topic: $ARGUMENTS (empty means the whole vault).

1. The cheap audit first

Read .claude/skills/vault-audit/SKILL.md and do its §§ 1–2 (run the tools, read the summary). Hold the report for the end.

2. Source independence

raw/_compiled/ may be deleted at any time (AGENTS.md § Source Independence). Reading it is legitimate here: the owner asked for this check.

For each compiled document — every file under raw/_compiled/, or those feeding the topic — pick three to five specific facts (a number, a name, a date, a table value) and answer each through the Query Protocol (_index.tsv route, one card, one _sections.tsv range) without opening raw/. A fact the wiki cannot answer is a compile gap: record the file, the fact, and the node where it belongs.

3. Judgement on watch items

For each watch item from § 1, say whether it is a real defect or correct as it is — thin people stubs, for example, are correct: padding them would mean inventing biography.

4. Gap audit

Tell the owner in one line how many agents will run and what they cover, then proceed. Run agents in waves of two, never more — wide runs hit the usage limit and lose everything. Split the topics (or the one topic) across agents. Dispatch each as the vault-gap-auditor agent — read-only, it skips the manual and carries the query ladder in its own instructions (D87) — and give it the vault path, its topics, and this brief:

Write 15 questions the owner would plausibly ask about <topics>, in their own voice — first person, casual, the way they type ("when did I…", "who's my manager on…"). Cover both work and personal material. Answer each only through the query ladder in your instructions: route on wiki/_index.tsv, pull one wiki/_cards.tsv row, then at most one wiki/_sections.tsv range — never read raw/, never read a whole node for one fact. For each question report: answered / partial / not answered; the trace (files and rows read); roughly how many tokens you read; and for any miss, the cause — a missing alias, a missing node, a missing fact, or a broken card.

After each wave, give a skeptic — another vault-gap-auditor — that wave's claimed misses:

For each claimed miss, retry the query with two other phrasings through the same protocol, and grep the wiki for the fact. Report CONFIRMED (with the fix: the alias to add, or the fact to capture and the node for it) or REFUTED (with the route that works).

Report only confirmed misses.

5. Report

The /vault-audit report shape, then three more sections: Only in raw/ (the compile gaps from § 2), Gap audit (confirmed misses with their fixes, and the answer rate — "27 of 30 answered, median trace ~250 tokens") and Watch items (the verdicts from § 3).

Add a scope line that says exactly what was tested: which topics, how many questions, how they split between work and personal material. Any coverage or fidelity figure states what it was tested against (D12) — Vifert's vault once claimed "100% fidelity" after testing only work-shaped entities.

End with "Which should I fix — numbers, "all", or "none"?"

6. Fix what they pick

As /vault-audit § 4. A routing miss gets its alias and its term goes into tools/probes.json, so the failure can never silently return (AGENTS.md § Query Protocol). A compile gap waits for the owner: capture it only if the fact is in the vault's own words elsewhere; if it needs raw/, list it for /vault-compile.