anthropics/synthesize

Synthesize

Gather

Never dump every finding into context. Get counts first, then pull what you need:

bun $ANT_CONTRACTS_CLI sql "SELECT kind, count(*) FROM findings WHERE run_id='$ANT_CONTRACTS_RUN_ID' GROUP BY kind"

For the per-doc tally, write a script (heredoc → bun) that reads findings and writes the projection to /tmp/tally.json — read THAT. Only SELECT … claim, quote … for the specific findings you're putting in the report.

Analyze with scripts, not in your head

For anything beyond reading — counts, joins, comparisons, tallies — write a short script (Bash heredoc → bun, or a scratch SQLite db under /tmp) and run it. The findings are already in SQL; most analysis is a SELECT. If you need a second Workflow fan-out for analysis, launch one.

Write the report

bun $ANT_CONTRACTS_CLI write reports '{"run_id":"'$ANT_CONTRACTS_RUN_ID'","brief_id":<b>,"body":"…markdown…"}'

Then one report_claims row per factual statement in the body, each linked via claim_citations. The v_uncited_claims view must be empty when you're done — check it.

Sampled semantic audit

When scope was 1 doc, 1 worker, and every report cite was minted via cli cite this session: the gate already enforced quote==span. Sample only the heading/subject context check; skip the byte-reverify. At corpus scale, do the full audit — it's the distractor-clause guard.

A real quote can support the wrong claim (e.g. a bare "§4.2.4.1.11: 5.5%" inserted by amendment has no topic — it's a savings percentage, not a withhold). Sample your report claims and, for each, re-read the cited span plus its surrounding section heading (and the base contract's parent section when the cite is a bare amendment insertion). Does the section's subject match the claim's subject? Is the cited document in the right family?

bun $ANT_CONTRACTS_CLI write audits '{"run_id":"'$ANT_CONTRACTS_RUN_ID'","kind":"semantic_sample","sample_n":<n>,"result":"<json or prose>"}'

A failed audit means the claim is wrong — fix the report and re-audit.

Recall sample

Check what scope excluded (computed, not something you set):

bun $ANT_CONTRACTS_CLI sql "SELECT * FROM v_scope_excluded WHERE run_id='$ANT_CONTRACTS_RUN_ID'"

If excluded == 0, write a trivial 0/0 audit row and stop — there's nothing to sample. If the question named a single contract by ID and reformulate's cross-ref grep already returned zero, the same applies.

Otherwise, full-read a random handful of the excluded set against the brief. Search for the entity names, agreement/reference numbers, section anchors, and defined terms you discovered DURING the sweep — not just the original scope predicate (re-applying the same predicate to the set it already excluded is circular and proves nothing). The point is to catch documents the scope filter missed because it didn't yet know the vocabulary. Book the miss rate:

bun $ANT_CONTRACTS_CLI write audits '{"run_id":"'$ANT_CONTRACTS_RUN_ID'","kind":"recall_sample","sample_n":<n>,"result":"{\"misses\":<m>,\"of\":<n>}"}'

Finish

set runs $ANT_CONTRACTS_RUN_ID status done. Then follow knowledge-harvest.