nx_osbench_judge.nx

nx_osbench_judge.nx -- rung F855 (lane osbench). The JUDGE: reads knowledge/status/osbench_measured.dat and COMPUTES every verdict from the numbers. It asserts nothing and it trusts nothing. Mirrors the proven nx_container_bench judge contract (measure elsewhere, grade here, refuse a different-config axis as an exceed) and adds the TIER rule that eats debt seq285. ★ THE THREE LIAR-KILLERS ARE NOT HYPOTHETICAL -- each one is a defect this lane actually shipped and had to catch by hand on 2026-07-20. The judge exists so the NEXT one is caught mechanically: L1 IMPLAUSIBLE a like-for-like store ratio beyond 10x either way is refused, not scored. The harness really did emit a 30-permille (33x) "win" that was pure leak-bias (Linux dir grew every run, NishiOS got a fresh image). This tooth catches that class WITHOUT needing anyone to notice a trend across runs. L2 CROSS-TIER only the sanctioned same-tier pair scores. An in-memory store compared against an on-disk one is durability-asymmetric (the original seq285 18x artifact). L3 UNSTABLE a ratio whose own cv exceeds the threshold is refused. Measured noise is not a result, and on a noisy host EVERYTHING must refuse -- that is the tooth working. ★ MISSION COVERAGE: weights live in knowledge/registry/osbench_weights.tsv (rule 11, data not code). The operator mission is RAW KERNEL PERFORMANCE, so context_switch/ipc/thread_spawn/syscall_trap carry 750 of 1000 -- and none of them is measurable yet. The judge therefore reports what share of the MISSION WEIGHT it could actually score, so a good number on the cheap axes can never be mistaken for a good number on the mission. An index without its coverage is a gap in a costume. VERBS: judge (default) | selftest (runs L1/L2/L3 against synthetic rows, in-process) license_tier: ORIGINAL No hw writes (Rule 26). expect_exit: 0