code wiki / _hdl_build / nx_graphics_board.nx
`nx_graphics_board.nx` source

↩ module page · 73 lines · 32305 B
1// nx_graphics_board.nx -- serves the sovereign graphics/game SOTA JSON (the /compare/graphics + MCP + atlas
2// interchange). Self-contained emitter of the GATE-PROVEN result from nx_graphics_sota_rollup (coverage 644,
3// 2026-07-18): the GATE liar-kills locally against real render artifacts; this BOARD serves the verified result
4// with NO file deps, so it runs anywhere -- including the NAS as an MCP tool. Re-bake when the gate score moves.
5// license_tier: ORIGINAL  expect_exit: 0
6import "nx_syscalls.nx"
7const K_MAGIC_262144: i64 = 262144
8
9func ocat(o: *u8, at: i64, s: *u8) -> i64 { var i: i64=0; var a: i64=at; while s[i]!=(0 as u8){o[a]=s[i]; a=a+1; i=i+1} return a }
10func aj(jb: *u8, jo: *i64, first: *i64, name: *u8, verdict: *u8, art: *u8) -> i64 {
  var j: i64 = jo[0]
  if first[0] == 0 { j = ocat(jb, j, "," as *u8) }
  first[0] = 0
  j = ocat(jb, j, "{\x22name\x22:\x22" as *u8); j = ocat(jb, j, name)
  j = ocat(jb, j, "\x22,\x22verdict\x22:\x22" as *u8); j = ocat(jb, j, verdict)
  j = ocat(jb, j, "\x22,\x22artifact\x22:\x22" as *u8)
  if (art as i64) != 0 { j = ocat(jb, j, art) }
  j = ocat(jb, j, "\x22}" as *u8)
  jo[0] = j
  return 0
21}
22
23func main() -> i64 {
  let jb: *u8 = sys_mmap(K_MAGIC_262144)
  let jo: *i64 = sys_mmap(16) as *i64
  let first: *i64 = sys_mmap(16) as *i64
  first[0] = 1
  var j: i64 = 0
  j = ocat(jb, j, "{\x22v\x22:1,\x22domain\x22:\x22graphics\x22,\x22title\x22:\x22Sovereign Graphics & Game Render (OS+Browser)\x22,\x22axes\x22:[" as *u8)
  jo[0] = j
  aj(jb,jo,first,"software-raster-core" as *u8,"HAVE" as *u8,"knowledge/nx_swgpu.png" as *u8)
  aj(jb,jo,first,"sdf-raymarch" as *u8,"HAVE" as *u8,"knowledge/nx_sdfbody.png" as *u8)
  aj(jb,jo,first,"trimesh-zbuffer" as *u8,"HAVE" as *u8,"knowledge/nx_trimesh.png" as *u8)
  aj(jb,jo,first,"smooth-shading" as *u8,"HAVE" as *u8,"knowledge/nx_phong.png" as *u8)
  aj(jb,jo,first,"pbr-material" as *u8,"HAVE" as *u8,"knowledge/nx_gx_pbr_ggx.png" as *u8)
  aj(jb,jo,first,"texture-mapping" as *u8,"HAVE" as *u8,"knowledge/nx_boxtex.png" as *u8)
  aj(jb,jo,first,"anti-aliasing" as *u8,"HAVE" as *u8,"knowledge/nx_aa.png" as *u8)
  aj(jb,jo,first,"global-illumination" as *u8,"PARTIAL" as *u8,"knowledge/nx_cornell.png" as *u8)
  aj(jb,jo,first,"skinning-animation" as *u8,"HAVE" as *u8,"knowledge/nishi_being_walk.glb" as *u8)
  aj(jb,jo,first,"asset-io" as *u8,"HAVE" as *u8,"knowledge/nx_bunny.png" as *u8)
  aj(jb,jo,first,"scene-graph" as *u8,"HAVE" as *u8,"knowledge/nx_scene_graph.txt" as *u8)
  aj(jb,jo,first,"parametric-subdiv" as *u8,"PARTIAL" as *u8,"knowledge/nx_bezier_patch.stl" as *u8)
  aj(jb,jo,first,"nishios-native-surface" as *u8,"HAVE" as *u8,"knowledge/nx_swgpu.png" as *u8)
  aj(jb,jo,first,"nishi-browser-3d-surface" as *u8,"HAVE" as *u8,"knowledge/nx_gx1b_page3d.png" as *u8)
  aj(jb,jo,first,"game-loop-canvas-surface" as *u8,"HAVE" as *u8,"knowledge/nx_gx1c_f0.png" as *u8)
  aj(jb,jo,first,"first-playable-game" as *u8,"HAVE" as *u8,"knowledge/nx_gx4_walk_t0.png" as *u8)
  aj(jb,jo,first,"game-entities-collision-winstate" as *u8,"HAVE" as *u8,"knowledge/nx_gx7_game_t0.png" as *u8)
  aj(jb,jo,first,"ambient-occlusion" as *u8,"HAVE" as *u8,"knowledge/nx_gx8_ssao.png" as *u8)
  aj(jb,jo,first,"save-load-persistence" as *u8,"HAVE" as *u8,"knowledge/nx_gx9_resumed.png" as *u8)
  aj(jb,jo,first,"mobile-portability" as *u8,"GAP" as *u8,0 as *u8)
  aj(jb,jo,first,"real-silicon-gpu-exec" as *u8,"GAP" as *u8,0 as *u8)
  aj(jb,jo,first,"cuda-benchmark-oracle" as *u8,"HAVE" as *u8,"knowledge/anat_explode.png" as *u8)
  aj(jb,jo,first,"digital-twin-anatomy-inout" as *u8,"HAVE" as *u8,"knowledge/sov_explode.png" as *u8)
  aj(jb,jo,first,"rigged-physics-body" as *u8,"HAVE" as *u8,"knowledge/rig_f0.png" as *u8)
  aj(jb,jo,first,"secondary-dynamics-softbody" as *u8,"PARTIAL" as *u8,"knowledge/rig_f0.png" as *u8)
  aj(jb,jo,first,"directx-compat-d3d11" as *u8,"PARTIAL" as *u8,"knowledge/warp_shaded.ppm" as *u8)
  aj(jb,jo,first,"tensor-core-neural-render" as *u8,"GAP" as *u8,0 as *u8)
  j = jo[0]
  j = ocat(jb, j, "],\x22coverage_permil\x22:827,\x22have\x22:20,\x22partial\x22:4,\x22gap\x22:3,\x22grounded\x22:27,\x22secondary_dynamics\x22:{\x22organ\x22:\x22nx_softdyn -- GENERAL spring-damper secondary-dynamics solver (soft-tissue jiggle, firearm recoil, hair, cloth) sovereign integer deterministic\x22,\x22gate\x22:\x227/7 GREEN: lag 6u, overshoot 40 sign-changes, energy 2.8k->0 rest@53, recoil peak 27 back@8, bounded 40/40, replay identical, ANTI-VACUITY stiff-lag 0 vs soft 6\x22,\x22caught\x22:\x22gate ran RED 2/7 first and found (1) a numerical-stability hole -- damping is a per-tick velocity multiplier, C above the envelope INVERTS+amplifies velocity -> envelope now ENFORCED inside the solver so an unstable config cannot be requested; (2) a measurement bug -- oscillation read in truncated integer units hid sub-unit wobble\x22,\x22honest_scope\x22:\x22solver PROVEN on physics; SOFT-TISSUE GEOMETRY NOW UNBLOCKED (GX-21: real 203k-tri body skin from the full BodyParts3D DB, same cadaver frame as the skeleton) -- remaining work is skin-weighting the surface to the solver points + tissue collision\x22},\x22soft_tissue\x22:{\x22source\x22:\x22BodyParts3D FULL whole-body DB (LSDB Archive), 1.26k OBJ meshes incl skin+muscles+viscera; CC Attribution-Share Alike 2.1 Japan per the FILE header (the web page says CC BY -- file is authoritative)\x22,\x22skin_mesh\x22:\x22FJ2810 = FMA7163 skin: 102k verts / 203k tris, sovereign-rendered 2.2s, overlays the skeleton with NO registration (same frame)\x22,\x22grade\x22:500,\x22grade_note\x22:\x22stats 550 SEMI-REAL coherence 997, perceptual ~500; up from 175 mannequin; judges AGREE = genuine climb\x22,\x22verified_by\x22:\x22nx_dr_run entailment judge BEFORE acting: true claim 809 covered, planted false claim 375 rejected\x22,\x22honest\x22:\x22geometry is now RIGHT, material is NOT -- uniform matte skin, no pores/SSS/hair, no eye shader, neck seam, blocky shadow\x22},\x22rig_physics\x22:{\x22organ\x22:\x22nx_anat_rig -- joint tree + FK + damped-pendulum integrator on the real 662k-tri skeleton, sovereign integer, no float/GPU/physics-lib\x22,\x22proven\x22:\x22posed->settle over 6 frames (shoulder 70->0 deg); REPLAY BIT-IDENTICAL on all 6 frame checksums = no float drift, no driver nondeterminism\x22,\x22honest_gaps\x22:\x22sagittal hinges only (no ball joints/twist); independent pendulums NOT a constrained rigid-body solver; no mass/inertia/collision/ground-contact; L+R limbs share a hinge; no muscle actuation\x22,\x22contact\x22:\x22GX-20 LANDED: root rigid-body under gravity + per-bone contact probes (lowest vertex, FK-transformed) + positional penetration resolve w/ restitution and rest threshold. Drop 260->15 w/ velocity accelerating then bounce+settle. NEG-CONTROL (contact disabled) falls THROUGH to -109 and diverging = non-vacuity proof.\x22,\x22next\x22:\x22ANGULAR root dynamics + per-bone collision shapes + friction so it can TOPPLE, be shoved and catch itself (today the root has only a vertical DOF, no angular momentum); then soft-tissue geometry to bind nx_softdyn to\x22}" as *u8)
  j = ocat(jb, j, ",\x22sub_censuses\x22:{\x22foundation\x22:688,\x22swgpu\x22:583,\x22sovgpu\x22:343}" as *u8)
  j = ocat(jb, j, ",\x22human_fidelity_permil\x22:175,\x22capability_level_permil\x22:175,\x22geometry_benchmark\x22:{\x22ref\x22:\x22REAL 15-part human SKELETON assembled from BodyParts3D (CC BY-SA 2.1 JP): ~662k tris, 15 layers, self-articulating from one shared cadaver frame, rendered on OUR sovereign renderer in 3.3s\x22,\x22skeleton_grade\x22:450,\x22skeleton_grade_note\x22:\x22stats 485 and perceptual ~450 AGREE -- real geometry moves BOTH judges, unlike the GX-13 FBM mottle that moved stats to 550 with perception flat at 140. Agreement = genuine climb, not Goodhart.\x22,\x22soft_tissue_human_still\x22:175,\x22vs_ours\x22:\x22real skeleton ~662k tris = 13.6x our WHOLE 48.7k-tri proxy body; ONE real skull = 3.5x it\x22,\x22known_source_gaps\x22:\x22Commons has NO radius/ulna (forearm absent, hand floats); sternum+cartilage occluded; bone material uniform (no porosity/SSS/AO)\x22,\x22route1\x22:\x22real CC anatomy = ORACLE ONLY (never the product), same posture as CUDA\x22,\x22route2\x22:\x22sovereign PROCEDURAL generation = THE PRODUCT, first byte up\x22,\x22PRODUCT_nx_body_gen\x22:\x22REGISTERED MCP TOOL. Our own body EMITTED from the anthropometric canon (per-mille of stature): each part is ONE continuous tube threaded through control rings and sub-sampled (no cylinder seams), with TRUE surface normals by central difference in both surface params, a 17-feature angular relief table, capped tube ends, and PER-PART PLACEMENT TRANSFORMS (canonical Y-tube rotated+offset -> real forward-running feet, and the mechanism fingers/toes will use); our own integer Bhaskara sine + our own IEEE-754 float32 ENCODER (NishiLang has no float type); ~26k tris, ZERO scanned asset, zero 3rd-party lib\x22,\x22MEASURED_vs_oracle\x22:\x22★★HONEST HEADLINE = 313 permil, NOT 837. A DETAIL judge was added to nx_bodybench because silhouette IoU (877/837) and front-view normal agreement (934) are BLIND to missing surface detail -- a smooth faceless mannequin overlaps a real human outline ~85pct. The detail judge measures local normal VARIANCE (surface busyness) as an IoU-in-variance-space vs the oracle: smooth-where-oracle-is-detailed AND noisy-where-oracle-is-smooth both punished (non-gameable by noise; the FBM Goodhart scores 0). It craters the headline to 313 = MIN(front 877, side 837, shape 934, DETAIL 313), matching the eye (operator: nowhere near the claimed level). True gradient: old v1 body=217 < current=313. ★★BODY IS NOW PROCEDURALLY GENERATED, not hand-carved: (1) nx_body_gen reads its whole canon (every ring + relief feature) from a DATA FILE, carries zero geometry literals; (2) nx_body_proc = a sovereign PROCEDURAL generator (Infinigen is the capability bar) that COMPUTES the canon from sex/build/musc/noise/seed/stylize via anthropometric dimorphism functions + seeded noise -- male/female/anime/individuals from ONE rule engine, zero hand-crafted/scanned asset. Pipeline params->generator->canon->emitter->mesh->honest-bench, every stage a registered sovereign MCP tool. Evidence sec22 https://nishifamily.com/world/anatomy_gpu. ★PROCEDURAL MUSCLE DEFINITION LANDED (sec23): the generator emits fine muscle anatomy BY RULES (loops) not hand-placed -- rectus six-pack GRID (2 cols x 3 rows + linea + tendinous grooves), serratus fan, deltoid heads, erectors, lats, biceps/quad/gastroc definition -- all a function of ONE muscularity knob (>35pct), scaled by it. MEASURED honest (per-region, not the whole-body average which barely moved 305->314 because abs=~5pct of pixels): torso-region surface busyness 4.62(soft)->7.20(ripped) vs oracle 7.91 = 91pct of the way, and APPROACHING not exceeding (7.20<7.91) so it matches real anatomy, NOT the FBM Goodhart. ★sec24 SOTA-ward pass: (1) DETAIL judge now MULTI-VIEW = mean(front yaw0, back yaw180) so it credits the whole surface not just the front (non-vacuity intact: oracle-vs-oracle 1000 on front/back/mean); (2) generated muscle definition extended to the LIMBS front-facing (rectus femoris/vastus/tibialis/forearm/clavicle) + back (erectors/traps/lats); (3) ★HANDS WITH FINGERS -- a digit-generation RULE places 4 fingers + thumb as tapered tubes from a flattened palm (first attempt splayed as planks, NOT shipped -- a bad hand is worse than an honest paddle; fixed by a real palm); mirror gives both hands, works across male/female/anime; (4) crotch-cap flap killed by a generated pelvic-floor ring. Honest headline ~344 (front-detail ~319, back ~369) up from 313 -- a REAL measured climb but still a ~1/3 surface. ★★★sec25 LAYERED FROM THE FIRST CELL UP (was: outside painting a shell). The emitter now builds the body as a NESTED LAYER STACK -- bone core (40pct radius, no relief), muscle envelope (82pct, full relief), skin (100pct + fat) -- each a real NXMSH2 LAYER that renders/peels separately (explode view shows bone|muscle|skin). The SKIN SHAPE EMERGES from the layers beneath, not painted on: a FAT param (subcutaneous, argv7) attenuates how much muscle relief reaches the skin, so lean shows the abs and heavy hides them FROM THE SAME MUSCLE LAYER = emergent, not a hack (proven A/B). Skin is outermost so it still benches vs the oracle (detail ~320-337, self 1000 intact). ⚠the bone layer is a scaled skeletal CORE VOLUME, not an anatomically-correct skeleton (no separate ribs/spine/pelvis); muscle is one envelope not named muscles -- next depth. ★★sec26 PER-FEATURE MATERIAL -> A FACE GETS EYES + CHARACTER CUSTOMIZER. Root fix for the faceless egg: every tri now carries its OWN material colour (written at emit), so a feature isn't stuck with the flat per-layer flesh -- eyes = white sclera + iris + pupil (iris colour = a material 1-4 brown/blue/green/amber, auto-varied from the seed). No renderer change (it already reads per-tri colour). Skin moved to LAYER 0 so the renderer's skin shading (wrap-diffuse) lands on it. Measured proportion fixes by fine-profile diff: legs SEPARATE (crotch was a merged block 228 vs oracle 119), shoulders raised to the acromion. ★CUSTOMIZER: one generator, params (sex/build/musc/noise/seed/stylize/eyecol) -> male/female/anime/builds/individuals, all with faces, ZERO hand-modelling. ⚠HONEST: still crude clay -- eyes small, no mouth/ears, head too egg-shaped, NO skin MATERIAL (SSS/tone) so it reads as clay; NOT photoreal. ★★sec27 SUBSURFACE SKIN LANDED in nx_anat_sov (the renderer/material wall): skin=layer0 now gets a real SSS term (red bleed rising as the surface turns from the key light -> warm/orange terminator glow instead of dead-grey) + multi-scale skin TONE variation (coarse blotches + fine grain). COLOUR-ONLY so it can't game the shape/detail judges. OCR-validated upward: torso before/after shows flat clay -> warm translucent flesh; the MATERIAL gap to the oracle has closed (oracle still wins on FORM). Renderer promoted (.prev-presss kept). sec28 MORE FACE: brows (two dark region-coloured bands, gap at nose = no unibrow), lips (lip-red region), bigger eyes, a projecting CHIN (lower face was RECEDING z62->30 -> curved into shadow; now steps forward to catch light), + a soft FRONT FILL light in the renderer (camera-facing skin no longer crushes to black). Face material = pMat 5 (colour-by-region on the head). Skin still benches self 1000/detail ~379. ⚠the mouth still reads too dark (faces DOWNWARD so little light reaches it -- a head-FORM limit not colour). ★★sec29 SPLINE LOFTING = the Infinigen-gap fix: the emitter interpolated LINEARLY between sparse rings (=stacked truncated cones + a crease per ring = the slab/cone look); now CATMULL-ROM cubic through each ring+neighbours (bg_cmr, integer-exact at rings, radius clamped >=0 for overshoot) = C1-smooth organic curvature from the SAME canon. Measured detail 381->389, self 1000. ★★DIGITAL-TWIN FRAME (operator): outside-in (oracle match) = the BENCHMARK; the GOAL is inside-out anatomical twin of a real being (human/horse/anything). Architecture already is that: INSIDE-OUT = the bone->muscle->skin layer stack (skin form EMERGES from muscle+fat); ANY BEING = canon-file-driven + species-agnostic spline engine (a horse = a different canon through the identical pipeline). ⚠HONEST: inner layers are still COARSE ENVELOPES not true anatomy (bone=scaled core not real ribs/spine/pelvis; muscle=one shell not named muscles; no organs); only 1 species (human) has a canon. ★sec30 SKIN MICRO-RELIEF: renderer perturbs the skin shading normal by a high-freq noise GRADIENT (2 octaves) over world pos = pore/wrinkle-scale relief that catches light -> surface stops being plastic. Renderer-only (mesh+bench untouched). SKIN MATERIAL STACK now 4 signals: Catmull-Rom spline form (sec29) + subsurface scattering (sec27) + tone variation (sec27) + micro-relief (sec30) -- all shading/colour, none touch the shape/detail judges. Surface quality now approaches Infinigen; the remaining Infinigen edge = FORM (real muscle striation/tendon/bone landmarks + higher mesh density) + inner-anatomy depth. Next: give each inner layer its REAL dense geometry + a 2nd species (horse) to prove generalization; wet-eye spec, hair/ears materials, face/form. NOT photoreal SOTA but surface is no longer clay/plastic. ⚠3-layer mesh: keep radial<=44 sub<=10 or it hits the 400k tri cap and truncates the skin head/hands (renders as red muscle). --- GEOMETRY-ONLY SILHOUETTE SUB-SCORE (blind to surface, kept for history): ★RULER IS NOW A SOVEREIGN REGISTERED MCP TOOL (nx_bodybench) -- the measurement left the dev laptop. It dual-renders ours+oracle to coverage+normal buffers on the NAS (height-fit framing, integer acos LUT, NO image round-trip, NO PowerShell, NO 3rd-party) and returns JSON in ~2s, so any agent/plan/workflow can measure a body with one call = THE LOOP SCALES. OOP: MeshView (one rendered mesh) + Judge hierarchy dispatched by kind (silhouette IoU, normal-field agreement) + BodyBench orchestrator publishing the MIN. On this sovereign ruler the current body = headline 837 permil (front_iou 877, SIDE_iou 836 = binding, shape 934), up from 834 after the HEAD became a real ring-tube part; PROVEN honest not gamed: self-compare 1000/1000, oracle-vs-oracle 1000/1000/1000 (measures identity on a DIFFERENT mesh, not saturated), and the OLD v1 body (pre-relief pre-foot) reads 760 vs the current 834 on the SAME instrument = it rewards real improvement on every axis. The retired laptop color-threshold ruler read 768 for this SAME body (stricter basis: it excluded shadowed-but-agreeing skin pixels); BOTH rulers agree the SIDE profile binds. shape=934 is high partly because front-view normals carry a camera-facing agreement prior -> headline is deliberately the MIN (bound by silhouettes), and a 3/4-view shape judge is the next ruler hardening. GX-24 climb (laptop basis) = 749->768 with FOUR results published that went the WRONG way or nowhere: (1) anatomical relief (17-feature angular radius table) moved SHAPE 749->749 EXACTLY ZERO -- correct refusal, relief modulates RADIUS while the normal still described the un-modulated ellipse, geometry grew bumps and orientation did not; (2) first true-normal fix scored 749->758 while the render came out BANDED like bandages -- one-sided in-band vertical tangent made the tilt constant per band and step at boundaries; rewritten (not patched) around one global station index with CENTRAL differences in BOTH surface parameters, continuous by construction -> 766 AND smooth; (3) capping the open tube ends removed a bright flat plate at the crotch (the camera was seeing the torso tube INNER WALL) and moved the metrics ZERO -- eyeball-only defect; (4) a REAL FOOT (per-part canonical-frame rotate+offset placement, Y-tube turned 90deg, flat sole by construction so ring-top=2*rb) REGRESSED the headline 766->750, and the regression exposed two real canon errors it had been hiding: mid-shank measured 78 permil deep vs oracle 36 (a 137mm calf, ~2x life size) and whole-body depth 126px vs 110 (torso too deep + over-long foot pushing the figure back in a height-normalised frame). Corrected -> 768, above the pre-foot best, feet kept. GX-24 climb 749->768 with FOUR results published that went the WRONG way or nowhere: (1) anatomical relief (17-feature angular radius table) moved SHAPE 749->749 EXACTLY ZERO -- correct refusal, relief modulates RADIUS while the normal still described the un-modulated ellipse, geometry grew bumps and orientation did not; (2) first true-normal fix scored 749->758 while the render came out BANDED like bandages -- one-sided in-band vertical tangent made the tilt constant per band and step at boundaries; rewritten (not patched) around one global station index with CENTRAL differences in BOTH surface parameters, continuous by construction -> 766 AND smooth; (3) capping the open tube ends removed a bright flat plate at the crotch (the camera was seeing the torso tube INNER WALL) and moved the metrics ZERO -- eyeball-only defect; (4) a REAL FOOT (per-part canonical-frame rotate+offset placement, Y-tube turned 90deg, flat sole by construction so ring-top=2*rb) REGRESSED the headline 766->750, and the regression exposed two real canon errors it had been hiding: mid-shank measured 78 permil deep vs oracle 36 (a 137mm calf, ~2x life size) and whole-body depth 126px vs 110 (torso too deep + over-long foot pushing the figure back in a height-normalised frame). Corrected -> 768, above the pre-foot best, feet kept. ★RULER CHECKED, NOT ASSUMED: oracle self-roughness (adjacent-pixel normal change) = 7.2deg vs ours 5.6, so the 36deg abdominal error was NOT oracle mesh noise; and a controlled A/B (relief scale 1000 vs 0, same code path) put the abdomen at 36.1 vs 36.2 deg, killing the relief hypothesis with a measurement instead of an argument. ★PER-PIXEL ERROR MAP beats band averages: the band average blamed the ABDOMEN, but those image rows also contain the arms/hands and the map shows the red is on ANATOMY WE DO NOT HAVE -- fingers, toes, face, deltoid-armpit. A horizontal image band is not an anatomical region. That map is now the to-do list, ranked by measured error rather than taste. ★AND THE JUDGE CAUGHT THE RULER CHEATING: after a depth edit the FRONT score collapsed 835->624 -- impossible from a depth change -- because auto-framing included width+depth, so a thicker body pushed the camera back and shrank on screen. The RULER was moving with the thing it measured. Fixed with height-normalised framing; front returned to 835, proving pure measurement artifact.  earlier: MIN(silhouette 834, SHAPE 758). Judge #2 is MECHANISTIC (no VLM): normal-field angular error 21.8 deg vs the oracle -- it immediately caught the silhouette metric over-reporting by 76 permil. Mechanistic judge family still to build: depth-field, curvature-statistic (MSDM2), Hausdorff, multi-view.  earlier: silhouette IoU 822 permil after a measured climb (v1 711 -> v2 668 REGRESSION -> v2-tuned 663 WORSE -> v3 arms-abducted 797 -> v4 shoulder+head+legs 822).  ★METRIC IS NECESSARY NOT SUFFICIENT: v4 scores higher but its shoulder reads as a BOXY SHELF -- silhouette IoU rewards WIDTH and is blind to SHAPE QUALITY, the same lesson as the statistical image grader that could not see fake texture. Next rung needs a shape-aware judge (curvature/normal agreement or the perceptual judge), not another width tweak. The two failures LOOKED better and scored worse; a per-band EXTENT diagnostic found the structural cause (oracle arm bands span x163-457, ours only 209-410 = real arms are held AWAY from the torso and end at hip level) and fixing that one fact moved +133 permil in a step.;  earlier baseline: anatomical width bands all within +-8pct (head +8, shoulder +5, chest -8, hip -3, thigh 0, calf -7) => the canon yields genuinely human proportions from parameters alone\x22,\x22honest_29pct\x22:\x22GX-24: real FEET landed and tube ends CAPPED; then the HEAD became a real anatomical ring-tube part (was a plain ellipsoid) -- deeper-than-wide skull + 9 facial relief features (nose/brow/chin/cheekbones/eye-sockets/occiput), inheriting true normals + caps. That surfaced a REAL BUG the big parts hid: station y was rounded to whole units BEFORE the finite-difference normal, so on a part 1/8 the body size the vertical tangent oscillated 6,8,6,8 -> horizontal BANDING (proven geometric by the normal map, immune to relief and to x/z precision because y never went through sub-units); fixed by sampling every coord at sub-unit precision. Still NO fingers, NO toes, NO eyes/lips/ears -- the per-pixel error map ranks those next, each buildable with the same ring-tube+relief+placement machinery\x22,\x22capability_benchmarks_legally_clean\x22:\x22BioDigital and similar platforms benchmarked on CAPABILITY only; we take NONE of their assets -- our geometry is our own procedural output or CC-licensed oracle data used as a ruler\x22,\x22INFINIGEN_BENCHMARK_VERIFIED\x22:\x22Read from Infinigen SOURCE (princeton-vl/infinigen, BSD-3, github). ARCHITECTURE: creatures = NURBS surfaces LOFTED through profiles along a central SKELETON CURVE (lofting.factorize_nurbs_handles; Length/Yaw/Rad polar surface coords; length_rad1_rad2 + aspect + angles_deg genes) -- SAME core idea as our ring-profile lofting, but (a) higher-order NURBS not our Catmull-Rom, (b) profile data captured from REAL references (nurbs_data/*.npy), (c) a many-PART GENOME (util/genome.py: body/head/leg/foot/tail/horn/hoof parts joined + smoothed, interpolatable between part types), (d) Blender Principled-BSDF material node groups per species (materials/creature/{tiger,giraffe,snake,fish_body,bone,eyeball,...}). Their SKELETON is a central curve for RIGGING/ANIMATION (util/rigging.py). ★KEY: Infinigen does NOT nest bone->muscle->fat->skin -- creatures are SINGLE-SURFACE. So our INSIDE-OUT LAYER STACK (real bone/muscle/fat/skin + subsurface visibility, veins/tendons through translucent skin) is a GENUINE DIFFERENTIATOR they do not have. Their EDGE over us: NURBS smoothness (we now have Catmull-Rom), profiles-from-real-refs, high mesh density, mature material nodes, and a full animal genome (we have 1 species).\x22,\x22HONEST_vs_REAL_PEOPLE\x22:\x22OCR'd our best render next to a REAL consented reference photo (private knowledge/, never republished): the gap is ENORMOUS -- real skin is EVEN + luminous with soft subsurface + soft-tissue form; ours is a crude mannequin. ★REGRESSION CAUGHT + FIXED: my coarse tone-variation (sec27) + micro-bump (sec30) read as DISEASED BLOTCHY MOTTLING at body scale (fine in a closeup, awful full-body) -- removed the coarse blotch, kept only a whisper of fine grain + even SSS. Real skin = EVEN, not mottled. NOT close to real people; surface-shading tricks != realism. The real climb = soft-tissue FORM (fat layer doing real work) + true inner anatomy + far higher density, NOT more shading noise.\x22,\x22evidence\x22:\x22https://nishifamily.com/world/anatomy_gpu (sec 7)\x22},\x22CEILING\x22:\x22GEOMETRY -- the mesh is a ~48.7k-tri stylized proxy (whole-body skin ~15.9k tris; entire nervous system ~2.6k tris for brain+cord+all nerves) vs ONE MetaHuman head ~100-160k tris. Organs are jellybean blobs, head is faceless, hands are mittens. NO shader fixes this; the real climb is high-density anatomically-correct GEOMETRY (real scan or procedural/Infinigen) + hardware for neural materials.\x22,\x22output_reality\x22:{\x22perceptual_grade\x22:175,\x22verdict\x22:\x22CLAY/STYLIZED -- anatomy-app mannequin on a floor; organs+outside look nothing like a real human (operator-confirmed)\x22,\x22ground_shadow_rung\x22:\x22landed 07-23: real projected contact shadow + per-pixel smooth normals + skin wrap/spec = no longer floating, less faceted (small real gain 140->175)\x22,\x22stats_grade_demoted\x22:550,\x22why_demoted\x22:\x22Goodhart: coherent FBM mottle moved scale+MSCN 309->550 with ZERO perceptual movement -- 2nd gaming vector after noise; stats grader = floor-filter only\x22,\x22headline_rule\x22:\x22capability = MIN(statistical, perceptual); perceptual = VLM/frontier-vision + operator eyes\x22,\x22real_photo_bar\x22:650,\x22photoreal_corpus\x22:776,\x22grader\x22:\x22nx_natstat ns_assess2 (floor-filter) + perceptual judge\x22,\x22measured_by\x22:\x22nx_grade battery + operator sight-check + realism loop /compare/graphics\x22}" as *u8)
  j = ocat(jb, j, ",\x22north_star\x22:{\x22external\x22:\x22MetaHumans (Epic) -- photoreal external digital humans\x22,\x22ours\x22:\x22IRL anatomically-correct being INSIDE AND OUT -- a HIGHER bar than MetaHumans (they do not render internals)\x22}" as *u8)
  j = ocat(jb, j, ",\x22photoreal_human_ladder\x22:[{\x22axis\x22:\x22skin-subsurface-BSSRDF\x22,\x22verdict\x22:\x22PARTIAL\x22,\x22artifact\x22:\x22knowledge/nx_gx_sss_skin.png\x22},{\x22axis\x22:\x22strand-hair-groom\x22,\x22verdict\x22:\x22GAP\x22},{\x22axis\x22:\x22eye-shader-cornea-refraction\x22,\x22verdict\x22:\x22PARTIAL\x22,\x22artifact\x22:\x22knowledge/nx_gx_eye.png\x22},{\x22axis\x22:\x22teeth-gums\x22,\x22verdict\x22:\x22GAP\x22},{\x22axis\x22:\x22facial-microgeometry-pores-wrinkles\x22,\x22verdict\x22:\x22GAP\x22},{\x22axis\x22:\x22scanned-material-maps\x22,\x22verdict\x22:\x22GAP\x22},{\x22axis\x22:\x22internal-anatomy-render\x22,\x22verdict\x22:\x22HAVE\x22,\x22artifact\x22:\x22knowledge/sov_explode.png\x22},{\x22axis\x22:\x22muscle-tendon-deformation\x22,\x22verdict\x22:\x22PARTIAL\x22}]" as *u8)
  j = ocat(jb, j, ",\x22sovereign_exceeds\x22:[\x22bit-exact deterministic render (pixel-exact vs Microsoft WARP)\x22,\x22one-owner stack lang-compiler-IR-raster-VM-OS-driver, zero libs\x22,\x22cross-surface: NishiOS + Nishi-Browser, no host WebGPU/WebGL\x22,\x22never-brick by construction\x22]" as *u8)
  j = ocat(jb, j, ",\x22sota_bar\x22:{\x22incumbents\x22:[\x22Unreal Engine 5\x22,\x22Unity\x22,\x22Godot\x22,\x22Bevy\x22,\x22three.js\x22,\x22WebGPU/wgpu\x22],\x22honest\x22:\x22CAPABILITY LEVEL = 140/1000 CLAY (PERCEPTUAL, operator-confirmed: plastic mannequin). Statistical grader said 550 but was GAMED TWICE (noise 475; then coherent FBM mottle 309->550 w/ zero perceptual change) -- demoted to floor-filter; headline = MIN(stats, perceptual). coverage 820 = BREADTH of primitives present -- NOT a capability level, do NOT read as SOTA-percent.  Old note: breadth != photoreal-human parity. OUTPUT REALITY: our human render is CLAY ~406 vs photo corpus 776 (gap ~370, realism loop organ-measured); human_fidelity ~200/1000 (SSS rung 1 + eye rung 2 shipped, gates+eyeball). FAR below MetaHumans (external photoreal digital humans) and FAR below our OWN inside-and-out anatomically-correct being (a HIGHER bar than MetaHumans). Presence != fidelity; the climb is SSS skin + strand hair + eye shaders + scanned materials + anatomy-correct internals.\x22}" as *u8)
  j = ocat(jb, j, ",\x22benchmark_oracle\x22:{\x22cuda_real_silicon\x22:\x22RTX 5080 compute 12.0, our kernel JIT->PTX, never-brick userspace\x22,\x22used_as\x22:\x22MEASURING STICK ONLY -- sovereign real-silicon-gpu-exec stays GAP (our own driver)\x22,\x22anatomy_twin\x22:\x228-layer scanned being 48728 tris rendered inside-and-out: all 8 systems non-zero in cutaway, stereo VR pair with measured disparity\x22,\x22evidence\x22:\x22https://nishifamily.com/world/anatomy_gpu\x22}" as *u8)
  j = ocat(jb, j, ",\x22keystone_gaps\x22:[\x22mobile agnostic-IR+aarch64 Gx-2\x22,\x22windows-linux-native-gpu-accel Gx-3 (RTX 5080, no FPGA)\x22]" as *u8)
  j = ocat(jb, j, ",\x22research_opportunities\x22:[\x22VM-executed programmable shaders (nx_wasm_vm as the shader core) - the unique sovereign move: determinism + one-owner, no incumbent has it\x22,\x22neural rendering via cooperative-vectors reusing our bit-exact HMMA/IMMA tensor-core encoder (serves DirectX neural + CUDA)\x22,\x223D Gaussian Splatting at SOTA (differentiable trained fit, not just a splat renderer)\x22,\x22Nanite-class virtualized micropolygon geometry (cluster LOD streaming)\x22,\x22Lumen-class real-time GI (path-traced bounce + IBL) beyond the Cornell radiosity PARTIAL\x22,\x22PBR image-based lighting (IBL) + scanned material maps -- the GGX/Cook-Torrance microfacet BRDF itself now SHIPPED (Gx-PBR: energy-conserving, integer, /world/gx_pbr.png)\x22,\x22sovereign real-silicon GPU exec (Path B: our DRM/Nouveau-class driver on the 5080, native not WSL-PV)\x22]" as *u8)
  j = ocat(jb, j, ",\x22fetch_targets\x22:[\x223D Gaussian Splatting real-time radiance fields (train + render SOTA)\x22,\x22Nanite virtualized geometry: cluster-LOD + micropolygon software rasterizer\x22,\x22Lumen real-time GI (software + hardware ray tracing)\x22,\x22cooperative vectors / neural shaders (DirectX SM6.9) + neural texture compression\x22,\x22mesh shaders + GPU work graphs (GPU-driven rendering)\x22,\x22ReSTIR spatiotemporal reservoir resampling (many-light path tracing)\x22,\x22DirectSR / DLSS / FSR neural super-resolution\x22,\x22OpenPBR + energy-conserving GGX BRDF reference\x22]}" as *u8)
  sys_write(1, jb, j)
  sys_write(1, "\n" as *u8, 1)
  return 0
73}
nx_graphics_board.nx source

`nx_graphics_board.nx` source
