Book Validate — Phase 01b: Layered Validation

Purpose

Prove that the converted Markdown is character-faithful to the original PDF/EPUB. Catch four classes of failure that downstream phases cannot recover from:

1. Dropped content — sidebars, footnotes, captions, alternate lines of long blockquotes.
2. OCR substitutions — wrong characters that pass spell-check (e.g., "rn" → "m").
3. Reordering — multi-column reading order broken.
4. Silent paraphrase — content rewriting that pdftotext/word-count cannot detect.

The harness is layered — cheap checks first, semantic checks last, LLM judge only on samples. Total cost for a 300-page book: under $1 in API spend, under 10 minutes of wall time.

Invocation

$ARGUMENTS should specify: