Closing the loop on metrics

Why it matters

After launch, operational signals have to turn into owned improvements rather than alerts nobody acts on. This lesson builds that loop while keeping every change behind a gate a person owns.

What to practise

Read the detection script: a rolling 30-day mean and standard deviation with Western Electric drift rules, deterministic and unit tested, with no model involved. Then read bands.yaml, where 1σ only logs, 2σ starts Claude read-only to diagnose, and 3σ may only open a pull request or run a pre-approved runbook. Follow the worked incident, a release that zeroes a cache, from breach to rollback to an agent-written intent.md. Watch where people stay in the path.

How it fits

bands.yaml is 6.5.2’s alert catalog in executable form: every tier has a first action, and a named service owner triages what the agent finds. The pre-approved rollback-deploy runbook is 6.5.3’s executable rollback for a release that went wrong. Triage into fix now, schedule or dismiss, and the eval each fix leaves behind, are 6.5.4’s use of operational signals to choose bounded improvements.