Expectrail makes game fixes prove themselves

Field notes

A fixed-route harness captures the changed game’s frames, then compares vision-model observations with the engine’s records using a deterministic evaluator. The page says uncertain or missing observations require review rather than counting as passes. In the described setup, the fixer can access its worktree and failed-check evidence, but not held-out bugs, private routes, evaluator code, check-writing access, or the release-signing key.