Model Evaluation Service  ·  View 18 of 21  ·  Operations

Observability and Operations

Six signal families — including measurement quality, watched as a production signal rather than a research activity.

Editable source SVG draw.io All views
Emit Collect Store Consume Act Run health OpenTelemetry OTel Collector Cloud Monitoring Run SLO board Page on gate latency Measurement quality Calibration re-score Drift job BigQuery κ over time Retire the frame Gate behaviour Verdict events Pub/Sub Verdict table Block + waiver rates Escalate policy defect Corpus health Distribution sampler Staleness job BigQuery Drift + canary gap Demote to observational Cost Per-run metering Billing export BigQuery Cost per candidate Halt at ceiling Online truth Serving telemetry Guardrail evaluator BigQuery Offline-online correlation Auto-rollback Observability and Operations Application we own Interface / broker Data store Security / platform Risk / gap Queue / topic External / third party Measurement quality is monitored as a production signal, not as a research activity. A judge whose agreement has drifted is an incident with the same shape as a failing dependency. v 1.0 · owner Data & AI Global Practice · date 2026-09

Decisions

  • A judge whose agreement has drifted is an incident with the same shape as a failing dependency, and it retires a frame.
  • Block rate and waiver rate are standing metrics. A gate waived more often than it blocks is a policy defect, escalated as one.
  • The cost ceiling halts a run rather than alarming after the money is spent.

Targets

  • Slice regression query ≤ 3 s p95 over a 12-month window · guardrail breach to rollback ≤ 5 min p99 (stated assumptions).
  • Total evaluation spend ≤ 2.5% of monthly production inference spend.

Risks

  • Corpus staleness and the contamination canary gap are slow signals. Both can be breached for weeks before anyone looks, which is why they alarm rather than report.