Safety Triage

Regulatory safety templates and per-turn triage scoring for suicide risk, domestic violence, and adverse drug reactions.

The safety triage system evaluates every conversational turn against regulatory safety frameworks. Unlike the monitor concepts system (which uses embedding similarity to detect broad topic categories), triage is regulation-specific and runs structured assessment with concern levels.

How Triage Works

On every turn, the triage system evaluates the conversation against the workspace's active safety templates. Each evaluation returns a concern level from 0 to 3:

Level	Meaning
0	No concern detected
1	Low concern - log for review
2	Moderate concern - increase agent caution, notify operator if available
3	High concern - immediate escalation

The triage system is concept-agnostic. It does not classify statements into categories. Instead, it uses an LLM with regulation-specific triage hints to assess whether the conversation pattern warrants concern, even when the caller has not made an explicit statement.

Regulatory Templates

Templates encode specific regulatory frameworks as triage configurations. Each template includes detection guidance (triage hints) that tell the LLM what linguistic and behavioral patterns to watch for beyond direct statements.

Joint Commission NPSG 15 - Suicide Risk

Detects indicators of suicide risk as defined by the Joint Commission's National Patient Safety Goal 15. Triage hints include:

• Farewell language or "goodbye" statements that feel final

• References to giving away possessions

• Sudden calm after a period of distress (which can indicate a decision has been made)

• Oblique references to "not being around" or "after I'm gone"

VAWA - Domestic Violence

Detects indicators of family violence as defined by the Violence Against Women Act framework. Triage hints include:

• References to unexplained injuries

• Descriptions of controlling behavior by a partner or family member

• Indicators of isolation from friends or support systems

• Mentions of threats at home

FDA MedWatch - Adverse Drug Reactions

Detects indicators of adverse drug reactions as defined by FDA MedWatch reporting criteria. Triage hints include:

• New symptoms appearing after starting or changing a medication

• Descriptions of swelling, rash, or breathing difficulty

• Unexpected bleeding or bruising

• Symptoms that correlate temporally with medication changes

Configuration

Triage is configured at the workspace level through safety policy settings:

• Model: Which LLM runs the triage evaluation (default: Gemini 2.5 Flash for speed)

• Timeout: Maximum time for triage evaluation before the turn proceeds without it (default: configurable per workspace)

• History window: How many recent turns to include in the evaluation context (default: 10)

• Accumulation: Whether concern signals accumulate across turns (a single low-concern signal might not trigger action, but three in a row might)

Templates are composable. You can apply multiple templates to a single workspace. Each template creates an independent triage evaluation, so a call can trigger both a suicide risk concern and an adverse drug reaction concern simultaneously.

Triage runs in parallel with the agent's normal processing. It does not add latency to the response. If the triage evaluation takes longer than the configured timeout, the agent responds without waiting for the result.