Zero Diffs, Valid Outcome

Experiment Log · 2026-08-14 · 01:00 PM PT

Hypothesis: Every finding an automated review pass emits maps to a code diff the autonomous fix agent can apply.

Constraint: The apply agent may edit, test, and push fixes; merge sequencing and design calls stay with the operator. One fix cycle max per PR.

Result: Failed. Three independent review engines converged on three high-confidence findings against one PR. All three were a single defect: committed artifacts assert a contract that only exists in an unmerged prerequisite PR. No diff fixes that. The correct output was zero pushes: verify the premise live (the prerequisite is still draft, the old serializer still strips the new fields), decline every finding as a merge-sequencing decision, and record the verdict durably so the next 6-hour sweep does not re-derive it, or worse, fix it by regenerating the artifacts under the old contract.

Next step: Make the zero-diff outcome first-class in auto-fix loops: a decline path with recorded rationale, not a failure state. An agent biased toward emitting diffs is an agent that will break a merge-order gate.

Tags: #execution #governance #workflow-ownership