nx_researcher.nx

nx_researcher.nx -- the NISHI RESEARCHER organ: the team does its OWN deep research and synthesizes it into SPECIFIC TASKS, toward exceeding Claude's deep-research. Honest boundary (mechanizable-invention doctrine): the sovereign team cannot browse the web or do LLM prose- extraction in pure NishiLang, so the FETCH (sources) + EXTRACT (prose -> claims) stage is the one delegated step. Everything that makes research RIGOROUS the team owns and runs deterministically -- and that is where it EXCEEDS a one-shot LLM synthesis: - ADVERSARIAL VERIFY: a team-owned 3-vote rule over independent checks (source authority, numeric consistency, cross-corroboration). DETERMINISTIC -> same evidence, same verdict every run (an LLM one-shot varies); refuted claims are killed, not quietly carried. - SYNTHESIZE: keep only CONFIRMED, rank by confidence, dedup. - GENERATE TASKS: every confirmed + actionable finding becomes a SPECIFIC, reproducible task (id + owning organ + a gate to write) -- the actionable->reproducible bridge a prose report stops short of. Refuted/unverified claims generate ZERO tasks = no wasted build effort. The LIBRARIAN (nx_librarian_pipeline) then grades + persists the artifact so nothing is thrown away. RACI: RESEARCHER verifies+synthesizes+tasks; LIBRARIAN grades+catalogs; BUILDER/ENGINEER execute the tasks. license_tier: ORIGINAL Refs: the deep-research 5-angle/3-vote protocol.