Sentinel

Evaluate, don't assert: why AI features need semantic monitoring, the three surfaces that share one judge (Sentinel in production, Workbook AI nodes, eval suites), how Sentinel works end to end, and the three design rules behind it.

Register an agent and get its key, send events to sentinel.verops.io or connect over MCP, the four rules that decide evaluation quality, the schema 2.0 envelope, the eight MCP tools, and how to verify the first run.

The twelve tabs — Control Plane, Runs, Explorer, Dashboards, Sentinel Health, Incidents, Behaviour, Actions, Monitors, Agents, MCP, Configuration — search, how to read a score, and a cookbook from a bad answer to a fix.

A run's states, the five base dimensions and the profile dimensions, how the independent judge works (consensus, grounding, provenance, never a zero), breach severity, incidents with RCA and recommendations, feedback and calibration, behaviour patterns and coaching.

The fifteen monitor kinds with preview and projected alert volume, cooldown, mute and ownership; the policy ladder from Observe to Scoped autonomous and the halt switch; the closed action catalogue; signed delivery through connectors; approving a recommendation.

The Explorer's dimensions and measures, saved views and CSV export; the six stock dashboards and the sentinel.* metrics; compare views that state a delta and refuse to over-interpret; the Sentinel Health page and the change digest.

The Workbook AI Evaluate node: the five criteria, thresholds, self-consistency, the deterministic safety backstop, and gating downstream nodes on the scores.

Golden datasets, the red-team preset, and gating deploys on AI quality.

A/B two models with an independent judge; the semantic knowledge base that grounds Workbook evaluations and Sentinel's grounding dimensions alike.

Runs that never appear, runs that never score, partial evaluations, 'judged without knowledge-base grounding', scores that moved when the judge did, monitors that never or always fire, approved actions that went nowhere, and MCP clients that see no tools.