nx_facemark.nx

nx_facemark.nx -- F1084 prereq 1: THE JUDGE THAT FAILS WRONG ANATOMY (beyond-metahuman program). WHY THIS EXISTS (measured, 2026-07-25): the detail judge scores LOCAL NORMAL VARIANCE and cannot ask whether that variance is anatomy, so a face carrying brow-flanges, a duckbill chin and no visible lips scored 357 against a cleaner face's 288. Three times now a variance-style judge has ranked wrong work above right work (FBM noise over a real render; a bell-shaped body over a good one; facial bumps over a clean face). A judge that counts busyness will always do this. So this one scores PLACEMENT. THE METHOD, purely mechanistic (no learned model, no rendering, no third-party): a human face read down its MIDLINE is a fixed sequence of alternating depth extrema -- brow(max) -> nasion(min) -> pronasale(max) -> subnasale(min) -> upper lip(max) -> mouth groove(min) -> lower lip(max) -> mentolabial(min) -> pogonion(max) Extract that profile straight from mesh geometry (per y-band, the most-forward vertex near the midline), find its extrema by prominence, and score ours against the oracle's on WHERE the features are. ANTI-SPAM IS THE POINT: the score multiplies recall (did we find the oracle's features, in the right places) by PRECISION (what fraction of OUR extrema correspond to a real one). A corrugated face full of bumps scores near zero because precision collapses -- which is exactly what the variance judge could not do. nx_facemark profile <mesh.nxmesh> -> the midline profile + detected extrema nx_facemark cmp <ours.nxmesh> <oracle.nxmesh> -> placement score + per-landmark table nx_facemark selftest -> gv gate license_tier: ORIGINAL expect_exit: 0 No hw writes (Rule 26).