nx_actbench.nx

nx_actbench.nx -- I3: THE PHYSICAL-PLAUSIBILITY RULER. (beyond-metahuman plan, Phase 0 instruments.) ★★★WHY THIS EXISTS, AND IT IS A MEASURED REASON RATHER THAN A PREFERENCE. The field published the number that convicts our whole instrument set: "Beyond MPJPE: A Physics-Based Audit of Monocular 3D Human Pose Estimation" (PhysHuman @ CVPR 2026) reports that the geometric standard MPJPE and the physical composite PHYSSCORE are correlated at only SPEARMAN rho = 0.37, WITH RANK REVERSALS -- the temporal method WHAM ranks 3rd on MPJPE and 1st on physical plausibility. Every ruler this program owns -- nx_twinbench (mm), nx_bodybench (geometry MIN), nx_facemark (placement), nx_curvebench -- is GEOMETRIC. So a body of ours can score well on all of them and still be physically impossible, and nothing we own could see it. This organ is the missing axis. ★WHAT IT IS NOT: invented teeth. The first draft of this rung was going to assert energy decay and replay determinism -- both fine properties, both MINE. runtime/nx_pose_benchmark_audit.nx exists precisely to convict that move, grading every claim MECHANICAL vs SELF-VALIDATED vs INDUSTRY-MEASURED, and its standing verdict is "ZERO industry benchmarks is the honest truth". So the channels below are the PUBLISHED ones (PHYSSCORE's six; definitions from the PhysDiff/EDGE lineage and PhyMotion arXiv:2605.14269), not the ones that were convenient to write. ★A RULER IS A FUNCTION OF A TRACE. PHYSSCORE operates on joint positions and needs NO GROUND TRUTH, which is the property that matters here: it can score a GENERATED being. That is what makes it usable on the creature/fantasy half, where there is no cadaver dragon to compare against. So this organ takes a trace + a skeleton and scores it. It is coupled to no generator. ★★CHANNEL 4 READS THE SKELETON'S OWN EMITTED TABLE. nx_skelgen writes "B <bone> <len> <lo> <hi>"; this reads those rows. There is therefore NO SECOND CONSTRAINT TABLE to drift from -- the exact failure nx_skelgen itself documents (a fix applied to one of two tables is a bug with a good alibi). ⚠★DECLARED SUBSTITUTION, because an undeclared one is a lie: PHYSSCORE measures penetration and float against the lowest MESH VERTEX. We measure against the lowest JOINT. A joint is inboard of the surface, so our penetration UNDER-reports and our float OVER-reports by roughly a limb radius. The number travels with that caveat in-band, every call. ⚠★AND THE FIELD ITSELF SAYS THERE IS NO STANDARD PARAMETER SET ("various methods may employ disparate parameter choices and even design distinct evaluation approaches"). So every threshold this organ uses is EMITTED IN THE OUTPUT. A plausibility number quoted without its thresholds is not comparable to anyone else's, including our own last run. nx_actbench score <trace> <skel.dat> nx_actbench selftest license_tier: ORIGINAL expect_exit: 0 No hw writes (Rule 26).