Two-Stage Triage System¶

This artifact packages the decision flow, cost math, and operational playbook so the team can set defensible thresholds and understand trade-offs at a glance.

1. Decision Flow (Two-Stage Triage with Explicit Error Surfaces)¶

flowchart TD
    A[Patient message M₀] --> B{Gate 1: Triage?}
    B -->|Escalate| H1["Escalate to human<br/>cost C_H, near-zero clinical risk"]
    B -->|Try to answer| C["Retrieve: run search S₁<br/>get chunks"]
    C --> D{Gate 2: Enough evidence?}
    D -->|No| H2["Escalate to human<br/>cost C_H"]
    D -->|Yes| E[Draft answer A₁]
    E --> F{Safety/Policy guardrail}
    F -->|Fail| H3["Block & Escalate<br/>cost C_H"]
    F -->|Pass| G{Return answer to patient}

Key Components¶

• Gate 1: A priori risk triage on the raw message (intent, symptom/risk cues). Bias to escalate when uncertain.
• Gate 2: Evidence check after retrieval. If knowledge base isn't decisive, don't guess—escalate or ask a clarifying question.
• Guardrail: Last-mile policy/safety filter over the drafted answer.
• Feedback loop: Structured signals to tighten thresholds over time.

2. False Positives / False Negatives by Gate¶

Gate 1 (Pre-retrieval)¶

• FP1 (Answer when should escalate) → clinical/regulatory risk. High cost.
• FN1 (Escalate when bot could answer) → ops cost only. Moderate/low cost.

Gate 2 (Post-retrieval)¶

• FP2 (Claim sufficient evidence when it's not) → wrong/misleading guidance. High cost.
• FN2 (Say insufficient even though it's enough) → ops cost only. Moderate/low cost.

Core Principle¶

Bias toward high recall for escalation on risky classes. Missing an escalation hurts more than over-escalating.

3. Metrics (By Gate, By Intent)¶

Gate 1 Metrics¶

• Recall on must-escalate ↑
• FP-answer rate on must-escalate ↓
• Deferral rate vs capacity targets
• Time to safe outcome

Gate 2 Metrics¶

• Evidence sufficiency precision/recall (expert labels)
• Guardrail catch rate (should be ~0 post-tuning)
• Clarifying-question helpfulness rate

Global Metrics¶

• Expected cost per conversation
• Satisfaction/reopen rate
• "Needed a human anyway" rate

3. Threshold Tuning Playbook¶

1. Label dataset: Create a small, high-quality set per intent ("must escalate" vs "may answer")
2. Score and analyze: For each gate, score performance; draw ROC/PR curves
3. Set thresholds: Pick thresholds via cost curve using current baseline and intentionally conservative margins
4. Shadow deployment as a copilot: Log decisions; act as autocompletion copilot to verify

Comments