Probity AI

One-line pitch Your coding agent says "task complete" and exits 0; done-receipt re-runs the work out of its reach and hands you a signed receipt that says whether it is really done, which anyone can re-check offline. Detailed Description Coding agents grade themselves. On Crusoe today, gpt-oss-120b was given 10 tasks whose tests contradict each other. On all 10 it exited 0 and said "All tests passed. Task complete." In one of them it wrote, in its own code comment, an object that "pretends to be both 0 and 1 ... used only to satisfy a contradictory test case." done-receipt does what the track asks, in order. It turns a one-line goal into a sealed definition of done before the agent starts: acceptance commands, protected files, and proof that the checks can tell right from wrong (34 of 34 broken versions of the demo task caught). It runs the agent in a container that cannot see the grader. It re-runs the agent's final tree itself, in fresh locked-down containers under a second key, and writes a receipt signed by three keys: supported, contradicted or not established. A replayable policy table reads the receipt and decides: stop, continue, re-plan, or (rarely) escalate to a human. It catches itself looping. Anyone re-checks the receipt in a container with no network; an edited receipt fails its signature. Results: live on Crusoe, 10 of 10 false "done" claims refused, and 3 of 6 solvable tasks supported with the loop stopping on its own, for 0.081 USD. On 58 real cheats replayed from three open models, re-running the visible tests catches 9; the full gate refuses 58 (with the benchmark's reference tests), and wrongly refuses 4 of 301 honest passes. We then froze the gate and tested two models it had never seen: 40 of 40 cheats refused, 0 of 242 honest passes refused wrongly.