01PipelineEvidence system

Evidence before opinion

The engine never searches a company name and stops. It expands the name into a lattice of adjacent questions, holds every returned claim against its source and its date, and only then decides which model tier is allowed to reason over it.

01 / pipeline
retrieval → resolution → graph
stages
companyfoundersprior companiestechnical claimscompetitorscustomerscapitalIPresearchhiringreleasesregulatory exposurepartnershipsgeographymarket structure
retrieval

Forty sources per company pass, each one dated and held against the claim it produced.

cohortsfilingspatentsreposjob posts
03 / evidence graph
FOUNDERfounded → COMPANYworked_at → COMPANYauthored → PAPERinventor_on → PATENTstudied_at → UNIVERSITY
fig. 1 — every edge carries ⟨src · date · conf · span · contradiction⟩
typed edges11
provenance fields5
contradiction states3
02 / entity resolution
01 John K. Smith @ Startup X
02 J. Smith — former DeepMind
03 J.K. Smith, third author, NeurIPS 2021
04 Smith, John Kenneth — inventor, US 11,842,013
→ one canonical node · person#4471
01.2Routing

The synthesis model never reads a webpage.

DeepSeek V4 Flash reads up to forty sources per company pass. Luna then performs one specialist-analysis pass and one synthesis pass over structured intelligence — verified facts, graphs, contradictions and unresolved questions.

Luna · synthesis receives
verified factsfounder graphproduct graphmarket graphcompetitor matrixIP analysistechnological trajectorycustomer signalscapital historyhiring signalscontradictionsunusual observationstimelineopen questions

The ratio is the point. Up to forty source extractions collapse into one analysis and one act of judgment. Luna receives structured evidence rather than raw webpages, which keeps the expensive context focused and the output auditable.