The Calibration Bench pairs a movable Baseline block with fixed Target slots for every criterion. Baselines come from real pre-pilot cycles — handling time, error rate, completion rate — not memory. Subjective-only baselines show low credibility until telemetry or a warehouse export backs them.
Escalations agents close without opening an engineering ticket
Credible · Prior pilot telemetryTarget sits far above Prior pilot — confirm stretch is intentional.
Median minutes from escalation open to resolution
Credible · Prior pilot telemetryTarget sits far above Prior pilot — confirm stretch is intentional.
Weekly escalation volume the workflow handles without backlog
Credible · Prior pilot telemetryTarget is reachable from Prior pilot without outlier weeks.
Escalation types explicitly out of scope for this pilot
Low credibility · subjective onlyTarget may be unmeasurable until Security reviews Allowed Exceptions.
Two of four baselines sit below Minimum — this pilot has to clear a bar the prior one didn't.
Saves data range, calculation method, exclusion rules, and evidence sources. After freeze, baselines cannot change without a recorded audit trail.