Indications and Warning Tripwire Watch
Stands watch on the indicator list from the running estimate and flags reporting that touches an indicator that has not yet been observed. Pushes to a human, never acts alone.
The exact use the IC names as generative AI best fit: standing watch and war gaming support, with the analyst owning the judgment. Assessment sections 4.1 and 6.
Human gate: Alerts push to a human. The agent never changes an indicator state or the assessment.
Trigger: On every reporting batch once wired (today: run manually).
- 01Read the indicator listRETRIEVEgetRunningEstimateAssessment
Indicators and warnings with their observed state, PIR, and NAI.
Success criteria- Indicator list read
Outputsindicators - 02Read recent reportingRETRIEVElistReporting
The reporting stream the tripwires are checked against.
Success criteria- Reporting read
Outputsreports - 03Check tripwiresREASONmatchIndicatorsToReporting
Match unobserved and partial indicators against report summaries; every alert cites the report that tripped it.
Success criteria- Every alert cites a report
- No indicator state changed by the agent
Outputsalerts - 04Analyst reviews alertsHUMAN GATES2 analystapprover: S2 analyst
The analyst decides whether an indicator state actually changed. The agent never updates the estimate itself.
Success criteria- Human review recorded
Outputsreview - 05Raise the alertNOTIFY (simulated)addTipCueMessage
Push reviewed alerts to the net and the running estimate owner.
Success criteria- Alert drafted
Outputsnotification
Wired mode runs per reporting batch and pushes alerts; dry run lists what would alert.
Prerequisites live in src/lib/agentic/client.remote.ts.template: published OpenAPI description, accepted Arazzo documents, a scoped server side credential, and an approved agent runtime. Until then, the selector stays on the local dry run engine.
Indicator records become first class objects the trigger watches.