Voice Register Audit

Voice Register 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

• During chapter rewrite under treatment class structural-polish or full-craft-rewrite.
• Whenever Wayne uses the Bucket 1 craft move "voice braid" from dev-docs/fiction-craft-rewrite-analysis.md.
• During pre-promotion audit to verify the chapter does not read as monolithic.

The audit does not apply to prose-polish or defamation-safe-tighten chapters — those classes do not commit to register braid.

The three registers

Tagging convention: each paragraph carries a leading HTML comment naming its register. Untagged paragraphs default to R-analytic (the book's home register), but a chapter using voice braid must tag the non-default paragraphs explicitly.

Decision rubric

A chapter passes voice-register-audit when:

1. Tagging coverage. At least 90% of substantive paragraphs (excluding section headings, list items, blockquote sources) carry a register tag. The 10% slack permits short transitional paragraphs.
2. Distribution within thresholds. No register is below 10% or above 70% of substantive paragraphs. (Thresholds are tunable from pilot data per rule 09; current defaults are the starting point.)
3. Transitions marked. Adjacent paragraphs in different registers are separated by an explicit transition cue — either a section break, an explicit attribution ("The inspector general's report states:" before an R-primary block), or a marked shift ("Stepping back from the inspector's account:" before an R-analytic block).
4. R-primary contents pass rule 06. Verbatim quotation discipline holds inside every R-primary paragraph.
5. R-frame contents pass rule 05. Newsroom-frame paragraphs use rule-05-compatible attribution verbs and hedging.
6. R-analytic contents pass rule 12 V8. The analytic voice does not commit internal laundering — it names structure rather than emoting at structure.

A chapter fails when:

• Tagging coverage is below the 90% threshold (the chapter is not actually braided; it is monolithic with sporadic flags).
• Any register exceeds 70% (monolithic by collapse) or falls below 10% (declared braid that does not happen).
• Register transitions appear silently between paragraphs — the reader does not feel the shift, so the braid does not pay its narrative cost.
• An R-primary paragraph contains paraphrase (rule-06 violation).
• An R-analytic paragraph contains emotional reasoning that exceeds the evidence (rule-12 V8 violation).

Conflict handling

1. Chapter contains long verbatim passage that pushes R-primary above 70%. Re-split the passage into shorter R-primary blocks interleaved with R-frame explanatory glue. If the source is genuinely so important it must dominate, document the exception in the audit output and route to xaiolai for confirmation that monolithic-R-primary is the intended voice choice.

2. Chapter has no R-primary at all. Permitted only when the chapter is genuinely about absence (no primary documents survive, or all are sealed). The chapter must explicitly say so. Otherwise route to Stephen to add primary-document quotes from the source ledger.

3. R-frame paragraph attributes a finding to a source whose evidence grade is C. Pass on register-audit terms, but flag for Stephen (rule 02 evidence-grade audit). Register-audit does not grade evidence; it grades distribution.

4. R-analytic paragraph contains the chapter's recognition beat (the hammer line). Pass — recognition is the analytical voice's defining moment. The audit verifies the recognition beat is in R-analytic, not R-frame (which would weaken it) and not R-primary (which would attribute the recognition to the source).

5. A paragraph carries two register tags. Hard fail; registers do not blend at the paragraph level. Either split the paragraph or pick one register. Mixed-register paragraphs were the failure mode that motivated tagging in the first place.

Escalation conditions

• Escalate to Wayne when distribution thresholds fail and the fix is paragraph rewriting.
• Escalate to Bonnie when distribution thresholds fail and the fix is scene reorder or new scene insertion.
• Escalate to Stephen when R-primary content fails rule 06 (quote integrity) or when R-frame content fails rule 02 (evidence grade).
• Escalate to Laura when R-analytic content fails rule 12 V8.
• Escalate to xaiolai when the chapter intentionally chooses monolithic register (e.g., entirely R-analytic by design) and the audit's distribution requirement should be waived for this chapter — waiver requires written rationale in the audit memo.

Boundary-case recipes

1. Auditing a chapter with three short scenes in three different registers. Per-scene register distribution can be 100% one register; the chapter-level distribution still has to satisfy the thresholds. If the three scenes are roughly equal length, distribution will be ~33/33/33 and the audit passes trivially.

2. Auditing the chapter opening. The opening scene's register choice signals the chapter's voice intent. R-primary opening (open on the document) is the project's preferred default per scene-construction/SKILL.md. R-frame opening (open on the inquiry) and R-analytic opening (open on the pattern) are permitted but Wayne should document why in the chapter brief.

3. Auditing the chapter ending. Beat 10 ("what this might mean for us") must be R-analytic. Beat 8 ("anti-laundering rule") must be R-analytic. If either ends in R-primary or R-frame, the chapter has lost its analytical voice at the load-bearing closing moments — flag.

4. Auditing a chapter whose voice braid was added by polish-only revision. Best-effort audit; the chapter's underlying structure may not support braid (e.g., no R-primary material in the source ledger). Recommend reclassification to prose-polish if the underlying support is absent.

5. Auditing a chapter that uses register-shifting within a paragraph for stylistic effect. Hard reject. The audit's premise is paragraph-level register tagging; intra-paragraph shifts are noise. Either split the paragraph or accept the chapter is not voice-braided.

Output format

Owner: voice-register-audit (skill run by <wayne|stephen|joe>)
Task: Audit voice-register distribution of book/chapters-v2/<n>-<slug>.md
Inputs reviewed:
  - book/chapters-v2/<n>-<slug>.md
  - book/chapters-v2/<n>-<slug>.contract.yml (for declared register intent if present)
  - .claude/rules/06-quote-integrity.md
  - .claude/rules/12-reader-experience-values.md
Output: voice-register-audit memo at process/audits/<n>-<slug>-voice-register-audit.md
Evidence grade: N/A
Assumptions: <e.g., paragraph counts exclude headings and list items per the rubric>
Open questions: <register-tagging gaps + recommendations>
Risks: <distribution drift if revision continues>
Handoff: <wayne to revise / bonnie to restructure / stephen for quote-integrity / nancy if R-frame attribution touches live actors>

Distribution summary:

Per-register findings:

Recognition-beat audit:

Examples

A: "Voice-register-audit for ch-07. Tagging coverage: 94% (3 untagged transitional paragraphs, within slack). Distribution: R-primary 24%, R-frame 31%, R-analytic 45%. All within thresholds. Transitions: 14 register-shifts identified; 13 marked, 1 silent at L290 (R-primary to R-analytic). Fix: add 'Set against the inspector's words:' before L290. Recognition beat at L580 in R-analytic — correct. Rule-06 spot-check on R-primary blocks: 8 verbatim, 0 paraphrase issues. Overall: conditional pass pending L290 transition marker." The skill identifies one minor issue (a silent register transition), recommends a one-line fix, and verifies the structural intent. Normal "conditional pass" output.