the story · the honesty machinery, graded
The Agent's Performance Review
The self-healing remediation agent gets the same treatment as the coaches: a public track record. Everything below is computed from the audit log the agent already writes — what it triaged, what it proposed, what the auto-merge gate decided, and whether each landed fix actually held for 14 days or the alarm re-fired. It is currently in shadow mode.
The record spanning 2026-05-29 → 2026-07-22
Did the fixes hold? — survival at 14 days
No fix has cleared the 14-day window yet — nothing gradeable, so no rate is claimed. A fix younger than 14 days is not-yet-gradeable — it is never counted as a success (ADR-104). “Regressed” means the same alarm class re-fired inside the window; “held” means the window elapsed clean.
Case files
- source freshnessPR opened for a human
fix(freshness): filter DDB query to DATE# SKs to prevent false-fire stale count (needs-review). Branch pushed: remediation/freshness-sk-sort-bug. PR creation blocked — GitHub Actions bot lacks createPullRequest permission on this repo. Operator must open the PR manually from the pushed branch. Root cause: two non-DATE# SKs (garmin REFRESH_RATELIMIT, food_delivery YEAR#2026) sort above DATE# in descending order, causing strptime ValueError and false-firing ~2 stale counts per day since ~2026-05-04. Fix: add begins_with(sk, 'DATE#') to the KeyConditionExpression in freshness_checker_lambda.py line 120. One-line change, no schema/IAM/logic impact.
source: agent-run log · 2026-05-31
view PR →remediation-log/2026/05/31/160107.json
Computed by scripts/v4_build_agent_review.py from remediation-log/ (the agent + auto-merge-gate audit log) via remediation/track_record.py · as of 2026-07-25
No new inference and no hand-kept numbers: this page reads records the platform wrote about itself. Case files pass an alarm-type allowlist before they render — security- or exploit-adjacent classes are excluded by design (R22), and the count withheld is shown above rather than hidden.