research-quorumlisted

# Research quorum Research has a property most quorum subjects lack: **checkable ground truth.** A citation either resolves or it does not. A number either appears in the cited source or it does not. That changes the shape. Per the corpus, generation with checkable ground truth is **Mine → Verify**, often with no judge panel at all — a panel over an uncontested check adds cost and no information. It also changes who mines: > **Route the mining to a research capability.** Do not run five agents doing overlapping web > searches. Judges score its verified output. --- ## The shape ``` 1. Delegate mining -> a research skill/agent that searches broadly and cites 2. Extract claims -> in code: pull every factual claim + its citation 3. Verify per claim -> refuters resolve the citation and check the claim against it 4. Judge (optional) -> only where the verified claims are contested or a decision hangs on them 5. Synthesize -> report with per-claim confidence, and what could not be verified ``` Step 2 is the one people skip. **A research report is not a finding; each claim inside it is.** The unit of verification is the claim, not the document. --- ## Extracting claims Turn prose into a checkable list before verifying: ```json [ {"claim": "NRR median compressed to ~101% in 2026", "citation": "data-mania.com/b2b-saas-benchmarks-2026", "type": "statistic"}, {"claim": "Expansion drives 40-50% of new ARR at scale", "citation": "abacum.ai/ru