audit

Audit (correctness pass)

Sweep the content of objective-subject notes for claims that are actually wrong, and hand back a confidence-gated findings list. This is the passive correctness layer: feynman and quiz only catch errors in notes you show up to engage with live; audit checks the notes you'd never think to re-open.

Core rules:

• Never write note prose. You diagnose: name the wrong claim, the correct version, and a source. The user repairs their own note. The only write you ever make is the audited stamp: a hidden <!-- audited: YYYY-MM-DD --> HTML comment appended at the end of the note, which you do by default (never prose); the user can opt out by saying so. It lives in a trailing comment, not a frontmatter key, deliberately, so it stays out of Obsidian's Properties panel.
• Objective matter only. Math, physics, CS, philosophy, etc. — claims with a defensible right answer. Skip personal/journal/opinion notes; if a target looks personal, say so and leave it alone.
• Web-verify suspect claims; never assert from memory. For any factual or attribution claim you suspect is wrong, search to confirm before surfacing it. If you can't confirm, drop it or mark it low-confidence — your own recall is not a source.
• Every finding carries claim → correction → source. No triple, no surface. "This looks wrong" without the correct version and a citation is noise; suppress it.

Read the vault profile (.vault-mind/profile.md) first if you haven't this session (paths + frontmatter schema); run /vault-mind:init if missing. Run from the vault root and substitute the profile's dir/key names below. The audited stamp is a trailing HTML comment, not a frontmatter key (see step 6); the profile documents it as such.

1. Resolve the target set

From the user's prompt, build the list of notes to audit:

• Topic ("my MSM notes", "philosophy of mind") → match notes by title/tags/body in <notes_dir>. When in doubt, list the matched notes and confirm scope before spending tokens.
• Tag (#physics, area/ml) → notes carrying that tag (respect the profile's tag style).

  rg -lN '^\s*-?\s*#?<tag>\b' <notes_dir>
  

• Folder → fd -e md . <folder>.
• Bare invocation / "since last time" → incremental mode (next step).

Pull in each target's wikilinked neighbors as context (so cross-note contradictions are visible) even though you only report on the target set itself.

2. Incremental skip

Audit stamps a trailing <!-- audited: <date> --> comment (see step 6). Skip any note already audited at or after its last edit, it hasn't changed since you last checked it. In bare/"since last time" mode this is the selector: audit every note whose audited stamp is missing or older than its file mtime.

# notes changed since their last audit stamp (mtime newer than the stamp, or no stamp)
for f in <notes_dir>/*.md; do
  stamp=$(rg -oNPm1 'audited:\s*\K[0-9]{4}-[0-9]{2}-[0-9]{2}' "$f")  # -P: \K needs PCRE2
  mtime=$(date -r "$f" +%F)
  { [ -z "$stamp" ] || [ "$stamp" \< "$mtime" ]; } && echo "$f"
done

Tell the user how many notes are in scope and roughly the cost before grinding through a large set; offer to narrow. Web verification is the expensive part, so the skip matters.

3. Load ground truth per note

For each target note: read the note and the source readings it cites (the profile's source-readings key, e.g. resources) so you're checking against what the note was built from, not just your own model. This is the same grounding move as /vault-mind:feynman.

4. Extract claims and sort candidate issues

Pull the note's checkable assertions (definitions, formulas, named results, attributions, causal/logical claims). Sort candidate problems into buckets, priority-ordered:

1. internal inconsistency — the note contradicts itself, or contradicts another note in scope/its neighbors. Highest signal, needs no web (it's pure cross-reference of your corpus). Cite both sides.
2. factual / notational error — wrong formula, sign, definition, constant, units, a flipped inequality, an off-by-one in an algorithm's complexity, etc.
3. misattribution — wrong person, date, paper, or who-said-what / who-proved-what.
4. conceptual conflation — two distinct ideas merged as if one (e.g. the Koopman operator vs. its generator; type vs. token; necessary vs. sufficient).
5. interpretive pushback (separate, low-priority) — not an error: an oversimplification, a missing caveat, or a claim you'd argue with. Walled off from the above so it never masquerades as "wrong."

5. Verify before surfacing

• For every candidate in buckets 2–4, web-search to confirm it's actually wrong and to get the correct version. Prefer authoritative sources (textbooks, primary papers, SEP for philosophy, official docs). Confirmed → keep, with the source. Can't confirm the note is wrong → drop it or downgrade to low confidence; do not flag on suspicion alone.
• Bucket 1 (inconsistency) is verified by quoting the two conflicting passages, no web needed, but if resolving which side is right needs a fact, verify that fact.
• Confidence-gate. Tag each survivor high / medium / low. Show high + medium by default; hold low-confidence ones back unless the user asks for the long tail.
• Watch for false positives that plague advanced material: deliberate simplifications, valid shorthand, context-dependent statements, and correct-but-unusual conventions are not errors. When the note is defensible under a reasonable reading, don't flag it.

6. Present, then stamp

1. Present findings grouped by note, errors first (buckets 1–4 in priority order) and the interpretive-pushback bucket last and visibly separate. For each finding give: file:line · bucket · confidence · the wrong claim → the correct version → source.
2. If a note came back clean, say so plainly, that's a real result, not a non-answer.
3. By default, stamp every note you actually checked (clean or flagged), no need to ask first. Skip the stamp only if the user opted out. The stamp is a single hidden line appended at the end of the note:

   <!-- audited: <today> -->
   

   If the note already has one, update its date in place rather than adding a second. Use an HTML comment, not a frontmatter key, on purpose: it stays out of Obsidian's Properties panel and is invisible in reading/live-preview. This is the only write. Do not edit note prose, even to "just fix" an obvious typo in a claim; that's the user's to fix (and mechanical frontmatter issues are /vault-mind:lint's job).
4. Offer to re-audit a note once they've revised it.

Scope guards

• Don't audit personal, journal, daily, or opinion notes; if the target set sweeps some in, name them and exclude them.
• Don't touch editor config, templates, saved-view, or asset files.
• Don't rewrite, merge, or delete notes. Findings in, metadata stamp out.