Set a rollback threshold from user-visible failures

CPU and error-rate alarms do not answer whether an AI release is harming users. Rollback criteria should use task outcomes and treat some events, such as unauthorized disclosure, as immediate triggers.

Define the measurable control

Choose a rolling window with minimum traffic. Define wrong high-impact outcomes, authorization failures, timeouts and safe refusals from the user’s perspective. Set thresholds before launch, map each to pause or rollback action and name the decision authority. Ensure the previous model, prompt and workflow version can actually be restored.

The AI reliability lab makes the decision reproducible with authored data:

from reliability_cases import rollback_threshold_case

result = rollback_threshold_case()
assert result["triggers"]["wrong_high_impact"] is True
assert result["triggers"]["unauthorized"] is False
assert result["triggers"]["timeouts"] is False
assert result["rollback"] is True

Five high-impact wrong results among 400 tasks exceed the one-percent fixture threshold. Timeouts remain below their threshold and no unauthorized result appears, yet one triggered condition is enough to roll back the candidate release.

Challenge the result

A tiny denominator makes percentages unstable, while a long window delays response. Combine absolute and relative conditions. Preserve new-version traces for investigation without continuing harmful traffic, and re-run the failed slice before promotion.

Keep a reviewable evidence pack:

• outcome taxonomy and measurement window
• pre-agreed rollback thresholds
• automated alert plus human authority
• rehearsed rollback and post-rollback verification

This work aligns with the evaluation, security, cloud operations, reliability and FinOps sequence in the FDE for Professionals course. The linked course describes the learning pathway; this article’s numbers are synthetic and do not report a model, client, audit or production result.

Practice task

Define three thresholds for a canary. Include one zero-tolerance condition, one rate with minimum volume and the evidence that confirms recovery.

Continue learning

This article is part of the Advanced AI reliability and assurance sequence. Use the neighbouring tasks when you need the prerequisite or the next application.

• Review the prerequisite or neighbouring task in Build a per-client AI cost attribution model.
• Continue with Run an AI incident tabletop exercise.

Reference: Google SRE Workbook: Canarying Releases.