anthropics/knowledge-harvest

Knowledge harvest

The knowledge index informs future reformulations. Auto-ratification is a feedback loop (one wrong fact biases every future brief that reads it), so you propose; a human ratifies.

When to skip

Skip harvest entirely when: (a) single-doc fact-lookup with a correct answer — there's no cross-doc structure to learn; or (b) the candidate fact is already verbatim in this run's brief or scope rationale (it's already recorded).

Check what's already there before proposing anything:

bun $ANT_CONTRACTS_CLI sql "SELECT fact FROM knowledge WHERE corpus=?"

Don't propose a near-duplicate.

What's worth proposing

Durable facts about the corpus that a future run would want to know during reformulation: "Ohio NextGen contracts use 'prompt pay' not 'clean claim' for the §4.2 timing clause"; "Acme amendments are cumulative, not replacing"; "the CCS framework's £100k cap is a framework-level figure, not per-call-off". Not findings about the question — those live in the report.

Propose

bun $ANT_CONTRACTS_CLI write knowledge '{"corpus":"<corpus>","fact":"…","source_run_id":"'$ANT_CONTRACTS_RUN_ID'","source_queue_item_id":<q or null>}'
bun $ANT_CONTRACTS_CLI write knowledge_citations '{"knowledge_id":<k>,"citation_id":<c>}'

Then surface it for ratification via the queue (the one human channel). The queue item's question is the fact itself, stated as a plain declarative — NOT wrapped in "Ratify …?" (that reads as a question about a question). State facts positively; avoid "does not imply" / double negatives. The ratification ask goes in context:

bun $ANT_CONTRACTS_CLI write queue_items '{"run_id":"'$ANT_CONTRACTS_RUN_ID'","brief_id":<b>,"round":<r>,"question":"<the fact, plainly>","context":"Proposed knowledge entry #<k> from this run — ratify or reject. Cites <doc.uri>.","blocking":0}'

Ratification happens in the chat front-end or inspector (writes ratified_by then status). The trigger refuses status=ratified without ratified_by. You never ratify yourself.