nx_referee.nx

nx_referee.nx -- the NISHI REFEREE / ADJUDICATOR (operator: "we seem to always get biased testing"; the racing role was scoped for running heats, but SCORING them fairly needs a separate, unbiased judge). The race in nx_race_vs_claude used CLAUDE'S OWN self-assessed accuracies -- the competitor graded itself, which is the exact bias to kill. The Referee enforces the bedrock of every fair judge system (code-golf / ICPC / Codeforces / ML leaderboards), bits-up: 1. SEPARATION -- the scorer is NOT a competitor (no self-grading) 2. GROUND-TRUTH -- score against a known-correct ORACLE, not opinion 3. HIDDEN HOLDOUT-- the test cases were not seen/tuned-on by the competitor (anti-overfit/Goodhart) 4. DETERMINISM -- same submission -> same score, every run (no LLM-judge variance, no flaky harness) 5. DECISIVENESS -- declare a winner ONLY if the margin beats a noise threshold; else TIE (no overclaim) A head-to-head is VALID only if all fairness gates pass. license_tier: ORIGINAL NOTE v1: core principles; to be EXTENDED by the running deep-research (contamination detection, adversarial/secret test generation, special-judge checkers, statistical-significance tests).