Signal Management
After Latitude discovers a signal, use the signal detail view to decide whether to investigate, monitor, resolve, or ignore it.Triage Workflow
When a new signal appears:- Review the description to understand the discovered failure pattern.
- Open example traces to see the conversations where it occurred.
- Assess severity: safety risk, quality degradation, or minor edge case.
- Choose an action:
- Generate an evaluation for ongoing monitoring.
- Resolve it if the problem is already fixed.
- Ignore it if it is noise or too rare to track.
Investigating a Signal
For signals that need deeper investigation, review several example traces and ask:- What user inputs trigger the signal?
- Is the agent consistently wrong, or is the failure intermittent?
- Are there shared patterns in context, tools, retrieval, model, or prompt behavior?
- If evaluations are linked, are they too strict, too lenient, or drifting from human review?
Resolving Signals
When the underlying problem is fixed:- Open the signal detail page.
- Click Resolve.
- Choose whether to keep monitoring.
- Enabled: Linked evaluations stay active and can move the signal to Regressed if the failure returns.
- Disabled: Linked evaluations are archived when the signal resolves.
Working with Regressed Signals
A regressed signal is a previously resolved problem that has returned. Treat it as high priority:- Review the new occurrence traces and confirm they match the original failure pattern.
- Compare them with the original fix and resolution notes.
- Fix the cause and resolve the signal again.
Ignoring Signals
Ignoring is different from resolving:- Ignoring immediately archives all linked evaluations, regardless of the keep-monitoring setting.
- Ignored signals are hidden from default views but can be restored later.
- Use ignore for noise, not for problems that have been fixed.
Keeping Signals Manageable
As your project matures:- Resolve fixed signals instead of leaving them open.
- Keep monitoring enabled when regressions matter.
- Ignore irrelevant signals so real problems stay visible.
- Keep signal descriptions clear enough for future teammates.
- Link meaningful signals to evaluations when you need ongoing coverage.
Next Steps
- Signals Overview: How signals are discovered
- Monitors: Get notified when signals are discovered, regress, or escalate
- Evaluations: Generate monitors from signals
- Annotations: Leave human feedback on signal traces