Dashboards and alerts

The control plane ships dashboards and alerting out of the box. Both are tenant-scoped and built on the same event stream the engine emits.

#Dashboards

A dashboard is a layout of cards: time-series, leaderboards, eval score histograms, cost-by-capability-class. Drag cards to rearrange, resize at the corners.

To create one:

1. Open the Dashboards page from the sidebar.
2. Click New dashboard.
3. Pick a layout template or start blank.
4. Add cards from the catalog. Each card binds to a query — pre-built queries are tenant-safe; custom queries are reviewed.

Dashboards persist to the dashboards table. The wire shape is documented under Dashboards in the API reference.

#Alerts

An alert is a rule that watches an event type, applies a filter, and fires when a threshold is breached.

alert:
  name: nightly-eval-regression
  source: eval.sweep_completed
  filter:
    suite: changelog-quality
  threshold:
    metric: rubric.mean
    op: lt
    value: 4.0
  channels:
    - integration:slack:ops

Channels reference integrations by id. The dispatcher fans the alert out best-effort — failures are logged and surfaced in the alert's detail view but don't tear down the run.

#Silencing

Alerts can be silenced individually for an interval (e.g. during a known-noisy migration). The silence is timestamped and audited.

#Test fire

Every alert has a Test button on its detail page that fires a synthesised event through the dispatcher. Use it to validate the channel wiring without waiting for a real regression.