Every prompt change the triage agent has proposed from reviewing its own outcomes. Each entry links to a draft pull request that a human reviews and merges — nothing here is applied automatically. This is the durable record of what the agent has learned, visible without GitHub access.
Each row can carry up to three separate verdicts, and they answer different questions.
The status is what a human did with the PR. The gate is what the before/after
eval concluded on labeled fixtures before the merge — and eval-unproven
is a real answer, not a failure: nothing regressed, but no fixture exercised the pattern,
so the eval declined to claim anything. prod is what actually happened to real
tickets afterwards. A change can pass the gate and still show
unverifiable in production, because only module routing and issue type have
a clean human signal to grade against; nothing on a ticket records why something broke.