Chapter Defect Diagnose

Chapter Defect Diagnose

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

• Before any chapter rewrite cycle is authorized. The defect map is the precondition for chapter-status: in-rewrite per rule 09.
• After a substantial revision that may have changed which values pass or fail.
• During Phase 0 of any book-wide rewrite project to build the per-chapter defect inventory.

This skill is the cited diagnosis that replaces "I want it more gripping." Without it, treatment classes are assigned on intuition and Gate B has no measurable starting point.

Decision rubric

For each of the 10 reader-experience values from .claude/rules/12-reader-experience-values.md, the diagnosis answers three questions with prose evidence:

1. Does the chapter pass this value as currently written? (pass | conditional | fail | unscored)
2. What specific paragraphs or sentences are the evidence for the verdict? (line anchors required)
3. What treatment, if any, is needed? (none | prose-polish | structural-polish | defamation-safe-tighten | full-craft-rewrite)

After diagnosing all 10 values, the skill picks ONE treatment class for the chapter per rule 08's table:

• All 10 values pass → no-change.
• Only craft values (V2/V4/V6/V9/V10) fail; core values pass → prose-polish.
• V7 or V8 fails on a live-content claim → defamation-safe-tighten.
• 1-2 craft + at most 1 core value fail; failure is structural → structural-polish.
• 2+ core values fail OR craft + core failures require Bucket 1 toolkit → full-craft-rewrite.

Conflict handling

1. Core value passes by craft but fails by evidence. V7 passes because the prose sounds calibrated, but the underlying claim is B-grade. The diagnosis records V7 as fail and notes "V7 fails on evidence-grade mismatch despite calibrated prose; treatment must lift the source, not the prose." This protects against the chapter sounding right while being structurally wrong.

2. Diagnosis output disagrees with prior audit. If the chapter has a current audit-chapter report that says PASS but defect-diagnose finds a core value failure, the diagnosis takes precedence (it is more recent and uses the rule-12 rubric explicitly). Flag the difference in the open-questions section.

3. Diagnosis cannot decide between two treatment classes. Pick the higher class. Class inflation is recoverable (rule 08 permits reclassification down with reason); class deflation hides real defects.

4. Diagnosis returns unscored on V6 (diagnostic memory) because the chapter has no beta-reader data. Score V6 against a structural proxy: does the chapter's anti-laundering rule (style-guide beat 8) name a transferable pattern in plain language? If yes, V6 is at least conditional. If no, V6 is fail. Do not leave unscored in the output.

Escalation conditions

• Escalate to xaiolai when the diagnosis assigns full-craft-rewrite to ch-01 or ch-13 — these are the book-boundary chapters and class assignment requires principal-author confirmation per the project's structural-stability principle.
• Escalate to Nancy when V7 or V8 fails and the chapter has live content (active litigation, named living individuals at chain positions, active corporate entities).
• Escalate to Laura when V1, V3, or V8 fails — Laura's red-team scope per rule 03 amendment covers exactly these values.
• Escalate to Bonnie when the diagnosis finds structural defects that imply the chapter brief itself needs revision (case selection, ordering, taxonomy match).
• Escalate to Stephen when V7 fails because of evidence-grade mismatch — fixing prose without fixing the source ledger merely papers over the issue.

Boundary-case recipes

1. Diagnosing a chapter at status: ready that has never been audited under rule 12. Run all 10 values from cold. Use prior audit-chapter reports for context but do not let them anchor verdicts. The rubric is new; the chapter has not been measured against it.

2. Diagnosing a chapter mid-rewrite cycle. Compare against the pre-rewrite snapshot in process/audits/history/<n>/<earliest-timestamp>/ rather than the original status: ready version. The diagnostic question is: has the rewrite improved or degraded each value?

3. Diagnosing a chapter whose case-file dossier is in flux. The dossier (book/evidence/case-files/) is the structural layer below the prose. If the dossier is being revised, defer defect-diagnose until the dossier stabilizes. Mid-flux diagnosis double-counts churn.

4. Diagnosing a chapter with documented partial-scapegoat taxonomy where V3 is structurally hard. Partial scapegoats (the visible actor IS culpable, but blame stops too low) make V3 (sympathy follows the chain) genuinely difficult; the reader's sympathy may legitimately split. In this case, V3 passes if the chapter installs a structural discrimination ("blame stopped here even though the chain extends further") rather than a sympathy transfer.

5. Diagnosing a chapter that has been rewritten and the snapshot is missing. The diagnosis flags the missing snapshot as a rule-09 violation. Do not proceed; either reconstruct the snapshot from git history or escalate to xaiolai.

Output format

Two artifacts per skill run:

Artifact 1: process/defect-map/<n>-<slug>.md

---
chapter: <n>
slug: <slug>
diagnosed_at: <ISO date>
diagnosed_by: chapter-defect-diagnose
chapter_status_at_diagnosis: <ready|in-rewrite|in-review>
---

# Defect Map: <chapter title>

## Summary

- Total values evaluated: 10
- Pass: <n>, Conditional: <n>, Fail: <n>
- Recommended treatment class: <one of the 5 classes>
- Confidence: <high|medium|low> with one-line rationale

## Per-value diagnosis

### V1 — Earned, not manufactured
- Verdict: <pass|conditional|fail>
- Evidence: <line anchors + 1-2 sentence quote/paraphrase>
- Treatment if fail: <one of the 5 classes>

[... repeat for V2 through V10 ...]

## Treatment class recommendation

- **Class:** <full-craft-rewrite | structural-polish | defamation-safe-tighten | prose-polish | no-change>
- **Rationale:** <which values drove the choice; why this class and not the adjacent one>
- **Downstream artifacts touched:** <list of review-memos, sidecars, case-file references the rewrite will invalidate>
- **Cross-chapter risk:** <which other chapters' callbacks / motifs / discriminations may be affected>

## Handoff

- Owner: <next agent>
- Open questions: <unresolved decisions for Gate A or principal author>

Artifact 2: append row to book/registries/treatment-classes.yml

Per the schema in rule 08. The skill builds the row idempotently — if the chapter already has a row, the skill appends to reclassifications: rather than overwriting.

Examples

assistant: "Diagnosis complete for ch-02 (the-four-goats). 6 pass, 1 conditional, 3 fail. V3 fails (sympathy lands on Calley by L312; chain at L380 is named but not weighted). V5 fails (partial/pure discrimination is announced at L210 but not consolidated; reader exits without the ability to apply it). V6 conditional (anti-laundering rule at L640 is portable but the chapter's memorable moments are scene-level not pattern-level). V8 conditional (one paragraph at L195 implies command knowledge without citation). Recommended class: full-craft-rewrite. Confidence: high. Downstream: invalidates 02-stephen-fact-check and 02-laura-red-team review memos; preserves all 9 case files. Escalating to xaiolai for class confirmation since ch-02 is in the forced reading-order chain (1→2→3) and a boundary-adjacent chapter." The diagnosis is specific (line anchors), cites evidence (V3 fail tied to L312/L380), routes per the escalation list (xaiolai because of structural position), and builds both artifacts. Confidence "high" because three failures all map to the rubric cleanly.