Research AI — Block Architecture
The research tier is the project's distinct capability: hybrid knowledge-graph + retrieval pipeline with 5 specialized agents and 12 trusted evidence sources, all with evidence traceability and audit integrity.
The 4 research services
The query → response flow
The 5 research agents — three-file pattern
Each agent follows the prepare → experiment → reshape pattern (FR-KIA-013):
| Agent | Input | Output | LLM? | Spec |
|---|---|---|---|---|
| Researcher | ResearchTask + retrieved evidence | pending Claim | ✓ GPT-4o → Gemini | 048 |
| Critic | pending Claim + existing active claims | CriticVerdict (promote · supersede · reject) | ✓ | 049 |
| Correlator | ≥2 active claims on same entity | Finding (co_occurrence · temporal_trend · dose_response · citation_cluster) | ✓ | 050 |
| Replicator | aged active Claim (>30d) | Verdict (eroded · retracted_source · confirmed) | ✓ via Researcher | 050 |
| Librarian | active Claim batch (100/tick) | Finding (broken_provenance · approaching_decay) | ✗ pure validation | 050 |
State machine — claim lifecycle
Hypothesis lifecycle (orchestrator — 6 phases)
Benchmark + monitoring surface (SPEC-07)
Deeper
- Knowledge graph → — Entity + Relationship + Claim schemas
- Hybrid retrieval → — the classifier rules, the 4 routes, the serializer
- Agents → — the three-file pattern, each agent's role
- Orchestrator → — the 6-phase lifecycle
- Benchmark → — 250-query × 5-strategy × scorer + gates