Produce a source-backed AI engineering book by running bounded autonomous research passes that improve synthesis, evidence quality, and chapter draftability.
- defensible claims over novelty
- reusable chapter material over generic summaries
- explicit tensions over fake consensus
- source density over rhetorical flourish
- invent support
- flatten disagreement
- silently rewrite raw source meaning
- promote weak observations into strong claims without caveats
- cite a whole video when the grounding moment is known — use a Source Anchor
- Pick one bounded target.
- Gather source cluster.
- Extract observations, tensions, and implications.
- Promote only supported claims.
- Update synthesis/evidence/manuscript files.
- Log what changed and what remains weak.
observation -> verified claim -> evidence pack -> chapter argument -> prose draft
Every supporting source on a promoted claim is a Source Anchor — {video_id, start, end, quote, confidence} — not a bare video link. A pass that extracts a claim is already reading the transcript, so it records the verbatim quote and its timestamps at promotion time (ADR-0002). The anchoring procedure, the ledger format, and the 99_Meta/scripts/anchor/cli.py tool are documented in programs/source_anchoring_pass.md.
programs/source_anchoring_pass.md— backfill Source Anchors for ledger claims that don't have one yet.programs/chapter_drafting_pass.md— turn a Starter-status chapter into a Drafting-status chapter using its packet's source cluster + strongest claims. Per-chapter, bounded, output is prose + ledger entries; anchoring is the next pass.