← Index

GATEDevtools & code

Agent 'Done' Claim Verification

This project verifies agent claims of task completion using Jev to prevent false 'done' signals.

Agents say "done." Sometimes it's fake. I use 0.5b-+ tokens/day on coding agents. I gate "done" with Jev @typesafeai before the PR, then scan sessions nightly. 7 days: 132 sessions, 65 "done" claims, 0 confirmed misses. Worst case: gpt-5.6-sol h/xh — 2/7 fake dones (29%).