Multi-Source Research Briefs: One Topic In, One Cited Brief Out

Ask a model to 'research X' in one shot and you get a fluent essay with the confidence of an expert and the sourcing of a rumor. The problem isn't the model's knowledge — it's the architecture: one trajectory, one perspective, no reconciliation. Real research is a fan-out, and it turns out you can automate the fan-out.

Fan-out / fan-in

How it works

1. Decompose the topic: the planner turns 'brief me on vendor X' into concrete sub-questions — pricing, funding, customers, recent changes, risks — each answerable independently.
2. Fan out in parallel: every sub-question runs against all three lanes at once; each lane returns findings as claims with source metadata, never prose.
3. Normalize claims: each finding becomes a typed record — statement, source URL or document ID, date, and lane — so reconciliation compares like with like.
4. Deduplicate across lanes: the same fact found on the web and in an internal memo merges into one claim with two citations — corroboration, recorded as such.
5. Detect conflicts: claims about the same entity-attribute pair with incompatible values (two different headcounts, two different prices) are flagged, not resolved silently; recency and source authority inform a lean, but both values ship in the brief.
6. Compose from claims only: the writing stage may only reference reconciled claim records — a sentence with no claim ID behind it fails validation and is dropped.
7. Report the gaps: sub-questions where no lane returned adequate evidence appear in a 'not established' section — the honest edge of the brief.

Reconciled claim record

{
  "claim_id": "clm_0192",
  "subject": "VendorX",
  "attribute": "pricing_model",
  "value": "per-seat, ~$40/user/mo",
  "citations": [
    { "lane": "web", "url": "https://vendorx.com/pricing", "seen": "2026-07-01" },
    { "lane": "internal", "doc_id": "doc_2211", "note": "sales call memo, May" }
  ],
  "conflict_with": "clm_0187",
  "conflict_note": "internal memo says $55/user - possibly legacy plan",
  "status": "contested"
}

Stack

• Orchestration — parallel branches with per-lane timeouts; slow lanes degrade to partial results rather than blocking the brief
• Web lane — search API + fetch + extraction to claims
• Internal lane — existing RAG index over company documents (see the document intake pipeline)
• Structured lane — read-only SQL against warehouse views
• Composition — LLM constrained to claim records; validation pass rejects unsourced sentences

A brief that says 'sources disagree on pricing: $40 (site) vs $55 (our May call notes)' is worth ten briefs that confidently pick one.