nx_dr_drbench.nx

nx_dr_drbench.nx -- SOVEREIGN DRBench + DEER deep-research SCORER / ruler (DR-0). Measurement-first: you cannot honestly claim "exceed Google" without the number. Implements the metrics the field actually scores, deterministically + integer-exact: DRBench (ServiceNow, arXiv:2510.00172): Insight Recall (THE binding constraint -- best system 36.5%, GPT-4o 13.2%), Distractor Avoidance (>93% everywhere = easy), Factuality (cited-claim support). DEER (LG AI, arXiv:2512.17776): Request Completeness + Evidence Validity (the two weak spots of every current agent), via back-tracking of cited AND uncited claims. The SCORER is sovereign/deterministic; the MATCHING (does a report insight match a groundtruth insight) is the separate judge stage (lexical/PPMI now via recall lane, no-float LLM escalation later) -- this organ takes the per-item match flags, exactly as nx_recall_eval takes graded gains. Imports ONLY nx_syscalls = drift-immune. No hardware writes (Rule 26). module: nishi-core.research.dr_drbench depends: nx_syscalls.nx genealogy_id: drbench_2026_servicenow + deer_2026_lgai NOTE: DISTINCT from nx_drbench.nx (a HotpotQA multi-hop QA harness) -- eats seq225.