Prompt Dependency Graph & Conflict Analysis¶

Date: 2026-04-23 Status: SUPERSEDED — historical reference only. This document maps the pre-PR-C composition stack where triage_base_v*.yaml was the base prompt and triage_agent.py:244-301 contained the routing logic. PR-C (#564, 2026-05-01) retired the triage_base_v* family entirely. The current composition is conversation_v4 (base) + phase_contexts/v2 + layer_contexts (additive), documented in docs/specs/conversation-v5-feature.md §1.3. Line numbers in this document refer to the pre-PR-C triage_agent.py and will not match current main.

For new readers: skip to the v5 spec for the canonical composition. This file is preserved for historical context on the conflict analysis that informed PR-B (additive composition) and PR-C (retirement).

Purpose (historical): Map how prompts composed at runtime and identify conflicting instructions before PR-B's additive-composition rewrite.

Composition Stack (Runtime Order)¶

┌──────────────────────────────────────────────────────────────┐
│  SYSTEM MESSAGE (sent to Claude Haiku 4.5)                   │
│                                                              │
│  ┌─────────────────────────────────────────────────────────┐ │
│  │  1. triage_base (triage_agent.py:244-301)               │ │
│  │     58 lines of inline rules: safety, tone, pacing,     │ │
│  │     documents, formatting, procedure-specific           │ │
│  │     ↓                                                   │ │
│  │  "Your current focus and guidance for this turn:"       │ │
│  │     ↓                                                   │ │
│  │  2. layer_context (config/prompts/layer_contexts/*.yaml)│ │
│  │     Active layer's turn-specific guidance + examples    │ │
│  │     ↓                                                   │ │
│  │  3. patient_context (injected at runtime)               │ │
│  │     Patient name, demographics, case info               │ │
│  │     ↓                                                   │ │
│  │  4. emotional_context (runtime)                         │ │
│  │     Detected emotional state of the patient             │ │
│  │     ↓                                                   │ │
│  │  5. forbidden_phrases (response_policy.py)              │ │
│  │     45 phrases the model must avoid                     │ │
│  │                                                         │ │
│  │  ─── CACHE BOUNDARY ───                                 │ │
│  │                                                         │ │
│  │  6. conversation_history (variable per turn)            │ │
│  │     All prior messages in the conversation              │ │
│  │     + [Clinical findings] block if docs analyzed        │ │
│  └─────────────────────────────────────────────────────────┘ │
│                                                              │
│  POST-GENERATION:                                            │
│  7. response_policy.check_response() — forbidden phrases     │
│  8. check_laterality_claims() — laterality guardrail         │
│  9. voice_rules.yaml — CI-time compliance scan               │
└──────────────────────────────────────────────────────────────┘

Key insight: Layer contexts appear AFTER triage_base in the system message. The model gives more weight to later instructions (recency bias). When layer context examples contradict base rules, the examples win.

All Instructions Inventory¶

triage_base (triage_agent.py:244-301)¶

intent_capture.yaml (Layer 1)¶

medical_status.yaml (Layer 2)¶

travel_readiness.yaml (Layer 3)¶

logistics.yaml (Layer 4)¶

financial_readiness.yaml (Layer 5)¶

Conflicts Found¶

CONFLICT 1: Question Count (CRITICAL)¶

B14 (triage_base:282):  "ONE question per turn. Never multiple."
  vs.
L1-3 (intent_capture:28): Example shows 2 questions in one turn
  vs.
L2-3 (medical_status:18): "Ask medications AND allergies together in one question"
L2-4 (medical_status:19): Example shows 2 questions combined

Winner at runtime: Layer context examples (L1-3, L2-4) — they appear later in the system message and the model follows examples over abstract rules.

Resolution: Fix layer context examples to show one question per turn. Remove "together in one question" instruction.

CONFLICT 2: Document Timing (CRITICAL)¶

B17 (triage_base:290):  "Never ask for documents until 3-4 turns in"
  vs.
L2-1 (medical_status:9): "TWO PRIORITIES: 1. Procedure 2. Documents"
L2-2 (medical_status:12): "Once procedure known, invite document uploads"

Procedure is typically identified by turn 2-3. So L2-2 says "ask for docs at turn 2-3" while B17 says "wait until turn 3-4."

Resolution: Change L2-2 to "Once procedure is known AND you've asked 1-2 follow-up questions about their situation, invite them to share records." This aligns with B17's spirit.

CONFLICT 3: First Turn Protocol (MEDIUM)¶

B15 (triage_base:284):  "FIRST message: ask how the PATIENT is doing"
  vs.
L1-3 (intent_capture:24-28): "Turn 1: Acknowledge procedure → clinical clarification → intent probe"

If patient states a procedure in their first message, B15 says ask about the patient's wellbeing, but L1-3 says acknowledge procedure + ask clarification.

Resolution: L1-3 should add: "If the patient mentions a person (child, parent, spouse) with a condition, ask how THEY are doing first (B15). Only skip to procedure acknowledgment if the patient is speaking about themselves."

CONFLICT 4: File Upload Messaging (LOW)¶

B9 (triage_base:269): "Say 'Based on your reports' — speak as one person"
  vs.
medical_status.yaml old text: "Our system is analyzing them" (was removed but pattern echoes)

Resolution: Already fixed — B9 is clear and the old text was removed.

CONFLICT 5: Efficiency vs Warmth (LOW)¶

B13 (triage_base:281): "Complete intake in ~10 minutes"
  vs.
B16 (triage_base:287): "Let the conversation breathe"

These aren't directly contradictory but create tension. The model may optimize for 10 minutes by reducing warmth.

Resolution: Already mitigated by removing turn-counting. Keep both — the 10-minute target is aspirational, "breathe" is the actual behavior.

Resolution Plan¶

Phase 1: Fix layer context conflicts (30 min)¶

config/prompts/layer_contexts/intent_capture.yaml: - Line 28: Split 2-question example into single question - Add: "Remember: ONE question per turn."

config/prompts/layer_contexts/medical_status.yaml: - Line 9-12: Change "TWO PRIORITIES" to sequential guidance - Line 18-20: Remove "together in one question" + split example - Add: "Remember: ONE question per turn."

config/prompts/layer_contexts/travel_readiness.yaml: - Add: "Remember: ONE question per turn."

config/prompts/layer_contexts/logistics.yaml: - Add: "Remember: ONE question per turn."

config/prompts/layer_contexts/financial_readiness.yaml: - Add: "Remember: ONE question per turn."

Phase 2: Live test (20 min)¶

3 conversations: 1. Mother + child leukaemia (Jeddah) — caregiver emotional 2. Patient + knee replacement — matter-of-fact
3. Patient + vague "options abroad" — exploratory

Verify: single question per turn, warmth, proper pacing, documents after rapport.

Phase 3: Monitor (ongoing)¶

Flag first 5 live conversations post-deploy. Compare against simulation baseline.