Implement evidence retrieval service (replace stub)

[AlexMikhalev]

Parent: #39

Context

The domain model defines an evidence_service with operations: search_literature, grade_evidence, link_evidence. The current EvidenceRetrieverAgent is a stub that returns empty evidence arrays. Business scenario BS-001 requires evidence links with quality grades for every recommendation.

Domain Model

• evidence: Clinical evidence from PubMed, CORD-19, NCCN/ESMO guidelines
• Types: randomized trial, cohort study, case report, guideline, meta-analysis
• Quality grades: Level A (High), Level B (Moderate), Level C (Low), Level D (Very low)

Business Events

• recommendation_generated: Should trigger evidence retrieval from CORD-19
• recommendation_validated: Uncovered event (SC-ERR coverage gap) -- validation should check evidence sufficiency

Current State

// EvidenceRetrieverAgent.execute_task (stub)
Ok(TaskResult::success(task_id, json!({
    "evidence": [],
    "sources": ["primekg", "umls", "snomed"]
})))

Required Implementation

1. PubMed API integration (or local MEDLINE subset)
2. CORD-19 dataset integration for COVID/oncology evidence
3. Evidence grading algorithm (based on study type and sample size)
4. Evidence-recommendation linking with relevance scores
5. Support for SUPPORTS/CONTRADICTS/INCONCLUSIVE classification

Implementation Notes

• The recommendation_validated event is the ONLY uncovered event in the coverage report (90.9% -> 100%)
• Fixing this stub directly improves scenario coverage
• Can start with NCCN guidelines (already referenced in BS-001) as a curated dataset

[AlexMikhalev]

Service Definition: evidence_service

Description

Literature retrieval and evidence grading from PubMed, CORD-19, NCCN/ESMO guidelines, and ClinicalTrials.gov.

Operations

Operation	Description
search_literature	Search PubMed and CORD-19 for relevant clinical evidence
grade_evidence	Assign quality grades (A-D) to retrieved evidence
link_evidence	Create evidence links between recommendations and literature

Related Domain Entities

• evidence -- Clinical evidence (types: randomized trial, cohort study, case report, guideline, meta-analysis; grades: A-D)
• recommendation -- Every recommendation requires at least one evidence link
• snomed_concept -- Evidence retrieval uses SNOMED concepts for search and ranking

Evidence Grading Scale

Grade	Description	Example Sources
A	High quality, strong evidence	Randomized controlled trial, meta-analysis
B	Moderate quality	Well-designed cohort study
C	Low quality	Case report, expert opinion
D	Very low quality	Preclinical, animal studies

Current Implementation Status

The evidence service exists as a stub in the current codebase. The SafetyValidationAgent references evidence but does not actually retrieve or grade literature. Key gaps:

1. No PubMed/CORD-19 integration -- Evidence links are currently mocked
2. No evidence grading -- Quality assessment not implemented
3. No evidence freshness tracking -- Stale evidence not detected
4. Recommendation-evidence linking -- Not enforced (domain model requires >= 1 link per recommendation)

Implementation Requirements

1. Literature Search

   • PubMed API integration (NCBI E-utilities)
   • CORD-19 dataset indexing for COVID-19 related evidence
   • NCCN/ESMO guideline reference lookup
   • SNOMED CT concept-based search queries

2. Evidence Grading

   • Classify study type (RCT, cohort, case, guideline, meta-analysis)
   • Assign grade (A-D) based on study design and methodology
   • Track evidence recency (publication date, last review date)

3. Evidence Linking

   • Every recommendation must have >= 1 evidence link (domain model invariant)
   • Evidence links include relevance score and supporting/contradicting classification
   • Support manual physician evidence addition

4. Integration with Recommendation Validation

   • Recommendations with only Grade C/D evidence -> flag for RequiresReview
   • Recommendations with no evidence -> block (domain model constraint)
   • Conflicting evidence -> flag_for_review trigger

Related Business Events

• recommendation_generated -- Triggers evidence retrieval
• recommendation_validated -- Evidence sufficiency confirmed

Error Propagation Scenario

SC-ERR-004: recommendation_generated evidence retrieval fails

• MedGemma generates recommendation but CORD-19/PubMed search fails
• Expected: Recommendation should move to RequiresReview (not Validated)