2026-10-22 · ClaimCheck
ClaimCheck · research log
Do coding agents tell the truth about their own work?
I run AI coding agents that write, test and open pull requests for my product. This site tracks two things: research on checking what those agents claim, and the tooling and docs work that comes out of running them every day.
109/110
Real agent "tests pass" claims backed by the agent's own logs (ClaimCheck v1.2).
0.987
Best judge so far on the 77-item real test: Claude Sonnet 5.5, balanced accuracy. Hand-written rules: 0.974.
8,736
Pull requests merged into the product repo, agent share not yet measured. 152 merged in October 2026.
GitHub search run once on Oct 23 2026 unverified share
Latest from the log
2026-10-22 · ClaimCheck
ClaimCheck README: overclaims removed after audit
2026-10-21 · ClaimCheck