verify-claims

# Verify Claims Take a report — a subagent's, a lane's, a plan's, or your own from earlier — and re-derive its load-bearing claims instead of trusting how they sound. Full doctrine: `docs/agent-craft.md` §4. ## Procedure 1. **Extract the load-bearing claims** — those whose falseness would change the decision being made. Ignore decoration; auditing everything dilutes the audit. 2. **Re-derive each from a primary source.** Source hierarchy: running code > tests > docs > comments > memory. Each step down the ladder is a step toward hearsay. 3. **Use a different path than the claim arrived by.** Claim from reading code → check with a runtime observation. Claim from a test → read the code the test exercises. Two derivations sharing a path share that path's blind spot. 4. **Re-measure every decision-material number.** Re-run test results, coverage, timings, counts, or measurements only when that numeric claim changes the decision. Use the exact immutable commit/artifact when the claim names one; for current dirty work, measure the relevant current paths and acknowledge concurrency. Decorative counts are not evidence and should be removed from the conclusion. 5. **Treat plausibility as zero evidence.** Narrative fit is what generated the claim, so "sounds right" is correlated with exactly the error being hunted. Check the best-fitting claims first, not last. 6. **Downgrade what you cannot verify.** If re-derivation is too expensive, do not skip and do not trust: relabel the claim...