nx_research_refute.nx

nx_research_refute.nx -- S2 of the Tesla ladder: ADVERSARIAL REFUTATION. Over a set of numeric claim-specs (a pattern the value follows, a claimed value in tenths, a tolerance) and fetched evidence text, it verdicts each claim (CONFIRMED / DISCREDITED / UNVERIFIABLE) and ADMITS ONLY CONFIRMED claims -- a claim survives only if the evidence AFFIRMS it; a wrong claim is DISCREDITED, an unfindable one is UNVERIFIABLE (never fabricated). This is the integrity layer FARS lacked (its own audit: 7.4% fabricated results, 5.7% hallucinated citations). Composes S1 (nx_research_synth_qwen produces the per-claim units to check) + nx_claim_verify_core (the verdict). The evidence FETCH stays in the caller (gate Tier-2 over sovereign TLS) so this organ is pure/deterministic/$0. license_tier: ORIGINAL