What this is

The shortest path to a real canonical output: prove one book end-to-end, read the clean store, then verify the registry. (Prove-one before batch — the engine's own operating rule.)

Prerequisites

• Node.js + ts-node (TypeScript, ES2020 / CommonJS). Install deps with npm install.
• LLM credentials in the environment (the pipeline routes through packages/core/src/llm.ts).
• The book's source asset present and registered (see registry commands below).

1. Run the book pipeline on one book

npx ts-node packages/collector/src/cli.ts <bookId>      # e.g. atomic_habits

This runs extraction → chapter detection → per-chapter summaries and writes under canonical_outputs/<bookId>/.

Operating note: chapter ingestion at scale runs off-local on Modal (the designed path), not the slow local lane — see the build-out docs. For a single prove-one, the local CLI is fine.

2. Read the canonical output (the clean store)

# clean, current book-level store — NOT the dead summary.json
cat canonical_outputs/<bookId>/book_level/book_level_sweep.json
cat canonical_outputs/<bookId>/book_level/_sweep_ledger.csv   # coverage

Per-chapter detail is under canonical_outputs/<bookId>/chapters/<chXX>/ (chunks.json, summary_core.json, gaps.json, summary.md).

3. Verify the registry

npm run registry:build       # rebuild from disk
npm run registry:verify      # confirm every asset exists
npm run storage:organize     # organize into ~/meta-factory-storage/

4. Honest verification (don't trust a too-fast run)

If a book "completes" suspiciously fast (e.g. 38s), it probably didn't run — confirm the outputs changed on disk before assuming success. Verify the unit, then scale to the batch.

See also

Concepts · Architecture · Trust & Provenance