Cite Density Audit

Cite-Density Audit

Global Five Over-Rules

1. Evidence before elegance. Never improve the story by weakening the evidence.
2. Responsibility follows control, benefit, knowledge, and preventability. Do not stop at the most visible actor.
3. Keep the taxonomy intact. Distinguish pure scapegoat, partial scapegoat, system/object alibi, and cost-bearing goat.
4. Steelman before judgment. Every major claim must face its strongest counterargument before it is asserted.
5. Handoff cleanly. Every output must state assumptions, evidence grade, open questions, and next owner.

When to use

• After Wayne or any agent edits a chapter prose file.
• Before any chapter promotion from in-review to ready.
• Across the corpus before running /compile-book (verbose brackets break the Chicago formatter).
• Whenever the source-ledger sweeps in changes that may have leaked metadata into chapter prose (the ch-06 Nielsen pattern that motivated rule 13).

The scan-cite-density.py hook fires on every edit at the pattern layer. This skill is the deeper read — verifies the corpus-wide invariant holds.

Decision rubric

A chapter passes cite-density audit when EVERY inline [CITE:] bracket satisfies ALL of:

1. Slug-only content. Bracket contains one or more card slugs separated by ;. Slug pattern: [a-z0-9]+(?:[-_][a-z0-9]+)*.
2. Length cap. Total bracket (including [CITE: prefix and ] suffix) ≤ 60 characters.
3. Card existence. Every slug resolves to an existing file at book/evidence/source-ledger/cards/<slug>.md.
4. No citation-apparatus phrasing. No commas, no quotation marks, no "reproduced verbatim in", no "tier-1", no "archived at", no "see also", no "Stephen verify" (the pending-stephen-lock suffix is the only permitted non-slug content per rule 13).

A chapter fails when:

• Any inline [CITE:] bracket carries a full citation ("Author, "Title," Journal vol. (year)").
• Any bracket carries a corroboration list ("AP, Reuters, NYT").
• Any bracket carries archive metadata ("archived at ...").
• Any slug points to a card that doesn't exist (broken reference).
• Any bracket exceeds 60 characters without justification.

Conflict handling

1. Bracket exceeds 60 chars but contains only valid slugs separated by ;. Permitted if 3+ valid slugs share one bracket (a load-bearing claim with multiple supporting sources can legitimately exceed the cap). Document the exception in the audit output.

2. Bracket contains pending-stephen-lock suffix. Permitted per rule 13. The suffix is the only non-slug content allowed inline. The audit confirms the slug exists and the suffix syntax is exact (<slug> — pending-stephen-lock).

3. Bracket contains a card slug that doesn't yet exist. Hard fail. Either Stephen creates the card OR the marker is converted to [EVIDENCE NEEDED: <description>] until the card lands.

4. Bracket contains a Layer-1 in-prose-source-naming pattern leaked into the bracket (e.g., the speaker name and date are inside [CITE:] instead of in the sentence). Hard fail; route to Wayne to restore Layer 1 in prose voice and reduce bracket to slug-only.

5. Bracket structure is correct but the surrounding paragraph lacks Layer-1 in-prose source naming (per rule 12 V10 and rule 13 Layer 1 requirement). Soft fail; route to Wayne to add in-prose source-naming. Cite-density audit reports the structural issue; the in-prose-naming pattern is enforced at audit-chapter / contract-audit time.

Escalation conditions

• Escalate to Wayne when prose needs in-voice source naming added to satisfy Layer 1.
• Escalate to Stephen when a referenced slug doesn't exist (card creation needed).
• Escalate to xaiolai when a chapter has 5+ over-length brackets — a structural sign that the chapter's source-naming pattern needs rework, not just per-bracket fixes.
• Escalate to jerry-crew-chief before running /compile-book — broken slugs or verbose brackets will cause Chicago formatter to fail or produce mangled endnotes.

Boundary-case recipes

1. Auditing a chapter with the legacy verbose-bracket convention. Expect many flags on the first pass. The chapter is pre-migration; normal. Wayne sweeps to convert each bracket to slug-only, moving citation metadata to the card (or verifying it's already there). Re-run audit after sweep.

2. Auditing a chapter with a [CITE: slug — pending-stephen-lock] marker. Verify slug exists and suffix syntax is exact. The pending suffix is permitted and signals deferred Stephen verification, not an audit failure.

3. Auditing a multi-slug bracket (e.g., 5 slugs separated by ;). Permitted if all slugs resolve and the load-bearing claim genuinely requires multiple sources. Length cap may be exceeded with reason documented.

4. Auditing a chapter against a corpus with deleted slugs. If a card was renamed or deleted in the source-ledger, the chapter's slug references break. Hard fail per "card existence" check. Route to joe + Stephen for ledger-sync.

5. Auditing during the migration sweep (Wayne dispatching per-chapter conversion). Acceptable to run cite-density-audit on an interim chapter state showing reduced (not zero) findings. The audit's job is to track progress, not to fail-stop mid-migration.

Output format

Owner: cite-density-audit (skill run by <joe|stephen|wayne>)
Task: Audit inline [CITE:] brackets in <path-or-corpus> against rule 13.
Inputs reviewed:
  - <chapter file(s)>
  - book/evidence/source-ledger/cards/ (for slug existence check)
  - .claude/rules/13-citation-form.md
Output: cite-density-audit memo at process/audits/<n>-<slug>-cite-density-audit.md
Evidence grade: N/A
Assumptions: <e.g., slug pattern is lowercase-kebab-case-with-digits>
Open questions: <slugs awaiting Stephen card creation; brackets requiring Wayne in-prose addition>
Risks: <chapters that would fail /compile-book due to broken slugs>
Handoff: <wayne for prose / stephen for card creation / joe for compile go/no-go>

Findings table:

Summary row: total brackets / valid / over-length / verbose-content / broken-slug / pending-lock.

Examples

A: "cite-density-audit for ch-06: 14 brackets scanned, 1 over-length + verbose-content at L31 (Nielsen tweet, 207 chars including 'reproduced verbatim in contemporaneous tier-1 reporting (AP, Reuters, NYT, WaPo, June 17-18, 2018)'); slug for card 'nielsen-tweet-no-policy-2018-06-17' exists. Fix: replace bracket with [CITE: nielsen-tweet-no-policy-2018-06-17]; verify Layer-1 in-prose source naming includes speaker, venue, date (current prose at L29 already names 'the Secretary of Homeland Security' but does not name Nielsen; add 'Kirstjen Nielsen' to the prose before the quote). Routes: Wayne for prose + bracket replacement." Real audit pattern: one over-length bracket caught by the scanner, slug verified to exist, Wayne gets specific fix instructions including Layer-1 addition. This is the Nielsen example the rule was written to prevent recurring.

A: "cite-density-audit corpus-wide: 287 brackets scanned. 198 valid (slug-only). 43 over-length but valid-content (multi-slug brackets in ch-04, ch-08 procedural-citation paragraphs; document as exceptions). 31 over-length AND verbose-content (legacy form, sweep candidates; concentrated in ch-02, ch-06, ch-08, ch-12). 12 broken-slug references (chapters cite slugs not yet in source-ledger; route to Stephen for card creation). 3 pending-stephen-lock markers (ch-09 L122 wd-sd-tex; ch-12 L93 walsh; ch-12 L19 black) — permitted per rule 13. Recommend: complete migration sweep on the 31 verbose-content brackets and create the 12 missing cards before /compile-book. Compile would currently fail on the 12 missing slugs." This is the corpus-wide pre-compile audit pattern. The skill produces an actionable manifest for jerry-crew-chief to triage before manuscript export.