Model analytics & monitoring

Every model on the platform, watched: live performance metrics computed from the Data Store and the audit trail (never self-reported), drift recognition against the population each model learned from, and alerting when a threshold breaks. An unwatched model is an unmanaged one — this page is the management surface's other half, beside the Control panel's registry.

Open alerts

Acknowledging is audited under your name; a condition that recurs after acknowledgment re-fires. Alerts never auto-close.

The registry, watched — every model

Complete coverage by construction: this table enumerates the model registry itself, so a model cannot exist without a monitoring row — including anything registered this session. Idle is a state, not a gap: a model with no invocations is watched and shown as idle.

Assistant-slot models share the slot's question volume (ai.query names the model serving each one); every other slot counts its own model.inference events. Deep metrics for the slots with performance surfaces follow below.

Denial risk model — performance & calibration

Calibration read: a healthy model's bands order cleanly — the high band's actual denial rate should sit well above the low band's.

Drift recognition — input population vs baseline

The other cores