Generated from .claude/skills/local/language-trap-agent-audit.md — do not edit here.

✎ Edit this page’s source

/language-trap-agent-audit — per-block agentic trap audit

The mechanical scanner (content/pipeline/language-trap-audit.ts) emits high-recall candidates. This skill is the required second layer: an agent reads each block in full and adjudicates all ten categories in context — including traps the regexes cannot see (a negation-contrast paraphrased without a comma, an aphorism that matches no template, a pivot in fresh wording). Owner directive 2026-08-15: every prose-bearing block gets this analysis; the script alone is not the audit.

The ten categories

Severity major for the diagnostic five (1, 2, 5, 6, 7 — rare in human scholarly drafting, near-universal in unedited model output); minor for the density four (3, 4, 8, 9) and the corroborating one (10), where the signal is concentration, not presence.

1. Negation-contrast — asserting by denying the opposite: “structural features, not afterthoughts”; “not X but Y”; “this is evolution, not a blank page”. A human drafter states the positive claim. Judge in context: mathematical contrast that is the block’s content is substantive (see false-positive classes).
2. Rhetorical pivot — setup-then-reframe: “that transformation is happening with or without a strategy; the question is whether it is governed.” The second clause carries everything.
3. Load-bearing em-dashes — the spaced dash substituting for sentence structure (appending afterthoughts, dodging the comma-vs-full-stop choice). Ranges and true parentheticals are fine; density is the tell.
4. Compulsive triples — the reflex three-item list regardless of underlying content. Substantive three-element enumerations pass.
5. Closing aphorism — a quotable-sounding final line placed because documents “end with quotable lines”.
6. Meta-commentary — narrating one’s own emphasis: “note the framing”, “we emphasize”, “I want to underline”, “the key takeaway”. Place the emphasis; do not announce it. Plain “note that” is standard mathematical prose and passes.
7. Thesis restatement — a block-final sentence restating what the text already established (“This establishes …” as a closer). “This completes the proof” passes.
8. Performed warmth — personal register in institutional prose: “thank you”, “we are excited”, “journey”.
9. Superlative inflation — maximum-strength claims without qualification: “most importantly”, “near-universal”, “single most”, “our mandate is clear”.
10. Non-speakable syntax — noun-heavy chains that parse on the page but stall aloud. Written for the eye, not the ear.

False-positive classes (record as pass with a note)

• Substantive mathematical contrast — the distinction is the claim: “a fixed property of the base, not of the fibre” in a block whose content is exactly that separation; “not only compact but Hilbert–Schmidt” (strengthening); “states, not modes” where the terms are technically distinct.
• Temporal status — “, not yet carried out” is a factual scope statement, not manufactured emphasis.
• Cross-reference closers — “This establishes clause 1 of the comparison summary” pointing at another block is bookkeeping, not restatement.
• Load-bearing disambiguation — emphasis that prevents a real misreading (e.g. a symbol collision) is kept, but stated directly rather than narrated (“This ‘fibre’ is distinct from the fibre functor $\tau$”, never “We emphasize that …”).

Per-block procedure

1. Read the block’s .md in full, plus title/kind from the .ts.
2. Judge all ten categories on the whole text — do not limit attention to the script scanner’s candidates.
3. Small confirmed traps (≤ 2 sentences, pure prose, no math and no meaning change): fix the .md directly. Keep the author’s content; remove only the construction.
4. Structural traps (rewrites beyond 2 sentences, meaning at stake, or math-adjacent): do not edit; record fail with a concrete proposed fix in notes.
5. After any edit, confirm math delimiters still balance: python3 -c "s=open('<file>').read(); assert s.count('\$\$')%2==0" and re-read the changed sentences aloud-style for register.

6. Record the verdict for every category with the helper — run AFTER fixes so the entry certifies the fixed state:

   bun run content/pipeline/qa-agent-entry.ts \
     --block <block.md> --criterion trap-negation-contrast \
     --result pass|fail [--severity major|minor] \
     --evidence "<what was examined / found>" \
     --notes "<fix applied: before -> after | proposed: ...>" \
     --id "agent:<batch-id>"
   

   Entries APPEND to the criterion array; script candidates stay in place as provenance.

Batch conventions

• Batches of ~25 blocks, grouped by chapter (mirrors qa-agent-drain-queue.ts). Disjoint file sets per concurrent agent.
• Agents in a shared tree do not commit; the coordinating session commits per completed batch: qa(trap-drain): <chapter> batch <n> — <fixed>/<pass>/<flagged> [bean].
• Each agent returns a summary: per-criterion counts, list of applied fixes (before → after), list of flagged structural items.

Relation to other audits

trap-* overlaps voice-ai-slop (category-H tells) and clarity-grad-style (register) but measures different constructions; verdicts are recorded under the trap-* criteria only. One-voice Category-H authorization applies: mechanical fixes need no per-edit author approval; structural rewrites do.