Audit Paper Book

A book companion goes stale the moment its source paper revises. This skill walks the gap between the paper and the book, classifies each drift item by mechanical-vs-substantive, and produces a single report. With --apply, the mechanical drift (new bib entries, new figures) is fixed in place; substantive drift always requires user judgement.

For NEW books, use /init-paper-book. This skill never creates a book that doesn't already exist.

Hard rules

1. Never edit chapter prose without explicit user approval. Mechanical fixes touch references.bib and figures/* only. Section content drift is reported, not fixed.

2. Numeric changes always block. If the paper changed a result number (mean gap, accuracy, theorem constant), the book reports it but does NOT auto-update. The user verifies the new number is intentional.

3. Atlas slug match. The book's slug must equal the atlas topic filename. Drift in either is a /init-project-research-level concern, not this skill's job.

4. Read-only is the default. --apply is opt-in; without it, this skill produces a report and changes nothing.

5. Accessibility floor. The book must remain readable by someone with an undergraduate degree in a quantitative field (linear algebra, probability, basic optimisation/statistics — not necessarily Bayesian methods or domain-specific machinery). Concretely:

   • Every acronym must be expanded at first use within each chapter (e.g. "expected hypervolume improvement (EHVI)").
   • Every display equation gets a one-line plain reading within 2 sentences.
   • No jargon-on-jargon sentences (≥3 specialised terms without inline definition).
   • Intuition before formal definition.

   The audit runs an accessibility check on every chapter (see Phase 2) and reports violations alongside other drift. Violations are reported in the accessibility drift bucket (see Phase 3), never auto-fixed. The user decides whether to revise prose or accept the violation.