Skip to main content
Signals is the queue of things that might need fixing. When a run executes, the Failure Agent detects problems such as ungrounded claims, tool-use mistakes, and guardrail blocks, then turns them into signals. Human 👎 feedback lands here too. Every signal starts as a candidate. You decide which ones are real failures — only signals you Confirm failure become fix targets and get grouped into a cluster. Signals you mark Not a failure drop out of the counts.

Screen layout

The Signals screen has two tabs.
  • Failures the list of detected signals. This page covers it.
  • Revalidation where accumulated beliefs get re-checked against a reference. See Revalidation.
The top right shows the total count (N signals).

The Failures list

The columns: You can filter above the list.
  • Search filters by summary text as you type.
  • Source filter All sources, or one of the A–G types
  • Flow filter All flows, or a specific flow

Working the queue

Click a row to open the detail drawer. Review the overview, detection evidence, and any human comment, then press [Confirm failure] or [Not a failure]. See Confirming failures. Confirmation is background: the drawer flips immediately and the cluster + case are created out-of-band. The signal row updates its verdict and cluster columns without a page refresh. Clusters are a separate menu in the sidebar; see Clusters.
Empty clusters don’t exist. A cluster is created only when a signal is confirmed and can’t be attached to an existing one — never from pending candidates alone.

Traces attached to signals

A trace with a live signal (candidate or confirmed) is protected from deletion — deleting it would erase the evidence the queue points to. The trace list and the builder’s run list surface a confirmation dialog when this happens; the delete only proceeds if you accept.

Next steps

Sources

The seven types (A–G) that produce signals.

Confirming failures

Review the signal detail and rule on it.

Clusters

Confirmed failures grouped by cause.

From signal to fix

The full journey from detection to deploy.