Evidence Research

Tool: evidence-research Model: Advanced

Trigger & Intent

Triggered by: Deep research, comparative analysis, or knowledge synthesis requests.

Intent: Triple-parallel synthesis using free lanes plus cross-model validation. Produces adversarially stress-tested evidence synthesis — not summarized opinion.

Resource Pooling

Capability profile: research — requires triple_parallel_synthesis, fan-out 3 (free lanes), then strong synthesis pass.

Pattern: GPT-5 mini (breadth) × GPT-4.1 (depth A) × GPT-4.1 (depth B, distinct framing) → Claude Sonnet 4.6 synthesis.

Required Skills

Skill	Role
synth-comparative-analysis	Side-by-side evidence comparison
synth-decision-engine	Evidence-to-decision translation
synth-recommendation-engine	Ranked recommendation synthesis
synth-research	Deep domain research

Input Schema

{
  researchQuery: string;
  depth: "surface" | "medium" | "deep";
}

Decisions & Throw-Backs

If 3 free lanes produce divergent synthesis (no majority agreement) → escalates immediately to enterprise for strategic framing before retry.

Success Chains

On successful completion chains to: plan · design · enterprise

FSM — Meaning-making with breakdown and reconstruction

stateDiagram-v2
    [*] --> ResearchQuery
    ResearchQuery --> TripleParallelFanOut

    TripleParallelFanOut --> FreeLaneA
    TripleParallelFanOut --> FreeLaneB
    TripleParallelFanOut --> FreeLaneC

    FreeLaneA --> PartialEvidenceA
    FreeLaneB --> PartialEvidenceB
    FreeLaneC --> PartialEvidenceC

    PartialEvidenceA --> Aggregation
    PartialEvidenceB --> Aggregation
    PartialEvidenceC --> Aggregation

    Aggregation --> CoherenceCheck
    CoherenceCheck --> SynthesisGate: majority agreement
    CoherenceCheck --> Breakdown: divergent outputs

    SynthesisGate --> StrongSynthesis
    StrongSynthesis --> EvidenceReport
    EvidenceReport --> [*]

    Breakdown --> Reframing
    Reframing --> TripleParallelFanOut

Execution Sequence

sequenceDiagram
    participant Orchestrator
    participant FreeLane A (GPT-5 mini)
    participant FreeLane B (GPT-4.1)
    participant FreeLane C (GPT-4.1)
    participant Synthesis (Sonnet)

    Orchestrator->>FreeLane A (GPT-5 mini): Broad research pass
    Orchestrator->>FreeLane B (GPT-4.1): Deep analysis pass
    Orchestrator->>FreeLane C (GPT-4.1): Alternative framing pass

    FreeLane A (GPT-5 mini)-->>Orchestrator: Perspective A
    FreeLane B (GPT-4.1)-->>Orchestrator: Perspective B
    FreeLane C (GPT-4.1)-->>Perspective C

    Orchestrator->>Synthesis (Sonnet): Synthesize 3 perspectives
    Synthesis (Sonnet)-->>Orchestrator: Evidence Report

    opt Divergent Outputs
        Orchestrator->>Orchestrator: Throw-back to enterprise for framing
    end