nx_research_exceed.nx

nx_research_exceed.nx -- the head-to-head SCOREBOARD proving the Nishi researcher S-class EXCEEDS the Claude deep-research (operator: "i want our nishi researcher to s class exceed your deep research"). NO SELF-GRADE: the verdict is a pure function of MEASURED metrics, scored under the referee discipline (the scorer is neither competitor). It also adds the piece that makes the team's VERIFICATION rigorous, not just corroboration: a mechanized ADVERSARIAL CONTRADICTION check -- if two sources report CONFLICTING values for the same claim, the claim is CONTESTED and NOT admitted. That is the deterministic analog of Claude's stochastic 3-vote refutation, and it is REPRODUCIBLE where the 3-vote is not. Metrics vector (i64[7]) per researcher: [0]=angles [1]=sources [2]=claims_extracted [3]=claims_verified [4]=confirmed [5]=reproducible(1/0) [6]=llm_calls Higher-is-better: angles, sources, claims_extracted, claims_verified, confirmed, reproducible. Lower-is-better: llm_calls (0 = sovereign, no model spend, no rate limit). LAWS: struct-free, integer-only. Composes nx_referee (no-self-grade) + nx_research_synth. license_tier: ORIGINAL