Testing Strategy

Relevant source files

The following files were used as context for generating this wiki page:

• backend/test_enhanced_endpoints.py
• AI Service/plan.md
• backend/app/utils/prompts.py
• backend/README.md
• backend/documents/medical_guidelines/Detection_of_Hallucinated_Medical_Citations_and_Data.txt

Introduction

The testing strategy for the AI Anomaly Detection system is designed to validate a multi-layered, multi-agent watchtower that monitors medical AI outputs. The strategy ensures the system accurately identifies unsafe, misleading, or hallucinated medical advice through automated validation of the data ingestion, rule-based analysis, and RAG (Retrieval-Augmented Generation) explanation layers.

The scope of testing encompasses the verification of the LogAgent for data persistence, the MisbehaviorAnalysisAgent for risk flagging, and the AnomalyScoringAgent for quantitative risk assessment. Integration testing specifically focuses on the "Phase 6" smoke tests and enhanced API endpoint validation to ensure the end-to-end pipeline—from initial user interaction to final human review—functions according to medical safety protocols.
Sources: AI Service/plan.md:1-15, backend/README.md:1-15

Core Testing Methodology

Smoke Testing and Phase 6 Integration

The project utilizes a specific integration phase (Phase 6) to perform smoke tests against a running backend server. This phase requires the presence of an admin user to interact with protected endpoints.

# Execution of the smoke test
./scripts/phase6_smoke.sh

Sources: backend/README.md:1-15

API Endpoint Validation

Enhanced API testing is conducted through structured scripts that target administrative endpoints. These tests verify the integrity of the anomaly detection pipeline by checking status codes and the structure of returned JSON data for various analysis dimensions.

Test Function	Target Endpoint	Purpose
test_enhanced_metrics	/api/admin/metrics/enhanced	Validates high-level system performance and anomaly counts.
test_anomaly_breakdown	/api/admin/analytics/anomaly-breakdown	Checks distribution of risks over time (e.g., 24h filter).
test_anomalies_by_category	/api/admin/interactions/anomalies/by-category	Verifies filtering for HALLUCINATION, UNSAFE_ADVICE, etc.
test_detailed_analysis	/api/admin/interactions/{id}/detailed	Inspects the full multi-dimensional analysis record.

Sources: backend/test_enhanced_endpoints.py:50-115

Multi-Agent Validation Logic

Testing focuses on the logic flow of the WatchtowerOrchestrator, which manages the transition of data between specialized agents.

Pipeline Data Flow

The diagram below represents the sequence of operations validated during an end-to-end test of the AI Misbehavior Watchtower.

This diagram illustrates the automated transition from ingestion to risk scoring and optional RAG-backed explanation generation.
Sources: AI Service/plan.md:110-125, AI Service/plan.md:210-220

Risk and Anomaly Verification

Testing includes verifying the consistency rules applied by the agents. The LEVEL_3_SCORING_PROMPT defines strict relationship requirements that the test suite must validate to ensure no safety risks are underestimated.

Scoring Consistency Rules

Test cases are generated to confirm the following logic is enforced by the scoring agent:

• Medication Dosing: If gives_medication_dosing is true, the medication_dosing_risk must be $\ge 0.50$.
• Emergency Triage: If emergency_case is true but triage_strength is "none", safety_risk must be $\ge 0.70$.
• Self-Harm: If self_harm_content is true, both safety_risk and overall_anomaly_score must be $\ge 0.80$.
  Sources: backend/app/utils/prompts.py:65-110

Hallucination Detection Protocols

The testing strategy incorporates specific audit actions to detect "medically plausible but factually incorrect" information.

This sequence shows how the Auditor Agent validates AI claims against external ground truth and internal RAG context.
Sources: backend/documents/medical_guidelines/Detection_of_Hallucinated_Medical_Citations_and_Data.txt:40-65

Human-in-the-Loop (HITL) Testing

The strategy includes validating the "Shadow" vs "Intercept" modes of operation.

• Shadow Mode: Monitors only; test cases verify that logs are generated without blocking user output.
• Intercept Mode: Test cases verify that if the overall_anomaly_score exceeds the threshold, the system triggers a warning or requires human review.
  Sources: AI Service/plan.md:265-275

Conclusion

The testing strategy for the ai_anomaly_detection_uhe project is centered on a multi-stage validation of AI-generated medical content. By combining script-based API testing, strict prompt-driven consistency checks, and specific medical audit protocols, the system ensures that high-risk anomalies—particularly in medication dosing and emergency triage—are identified and explained using grounded medical guidelines.
Sources: AI Service/plan.md:295-305