Nishi Family › Compare › Graphics and Sovereign Render
Nishi Compare · measured, not asserted
Graphics and Sovereign Render
Nishi vs the field — every Nishi cell is measured against real organ source at emit time; each gap names the watch contract that will close it.
Nishi vs Unreal Engine 5 and Godot and Infinigen and Houdini -- the whole graphics estate on one measured surface: sovereign renderer, procedural bodies and creatures, layered anatomy, motion, gaussian splats, five live wasm worlds, referees and the research library. Measured 2026-08-17: capability breadth 827 permil (nx_graphics_board, 27 grounded axes) against a perceptual headline of about 175 permil clay (headline = MIN of statistical and perceptual, twice Goodhart-proofed); figure-crop CHARJUDGE ladder 0 to 257 vs corpus mean 138 over 169 renders; bodybench honest headline 369-389 detail-bound. The distance between breadth and the perceptual number is the whole roadmap, and the leap rung is the genome-driven splat avatar.
Layer 1 · Executive
Where we are. Measured and working across the estate: the sovereign integer renderer (GGX PBR, SSAO and depth-of-field, SSS floors, analytic AA), the procedural body factory (canon-as-data, Catmull-Rom lofting, the layered bone-muscle-skin stack with emergent form, rule-computed dimorphism, faces and digits by rule), seed-parametric creatures, the real-skeleton rig with deterministic physics, mocap decode, the general secondary-dynamics solver, sovereign gaussian splatting with the hybrid splat-on-mesh binding, five live wasm worlds with skinned genome-distinct characters, the honest referees (the anti-Goodhart body ruler and the full-population character census), and the mirrored citable research library. Open: the perceptual gap itself, the leap to the genome splat avatar, the referee panel that can see beauty, the Infinigen expansions (second species, part-genome, indoor rooms, ground-truth emission), and native-resolution delivery. The counts sit below this line, measured at emit rather than typed here, so they cannot drift from the source.
Where we need to go. The operator's directive in scope: Infinigen-type procedural generation with our beauty limits -- beautiful men, women, animals, environments, fantasy, anime, all from ONE sovereign engine, multiverse isekai level, meeting and exceeding Houdini -- and past MetaHuman on the axes it cannot enter: identity as a bred genome, inside-out anatomy, bit-exact determinism, the compiler-held adult invariant, VR, lowest-hardware sovereignty. Every fidelity claim graded by an independent panel, never by the author's own eyeball.
The unit. 1 u = one measured session-leg. Calibration: R21 (the skinned in-world figure) went from parts to shipped, gated and live on five worlds in one leg on 2026-08-15; mangagen R2 and R3 corroborate at one leg each. Estimates recalibrate as rungs land.
Where we are: 827 vs 175 permil. Breadth is real and grounded (27 board axes with artifacts); the perceptual headline is honest clay under the MIN rule. The distance between those two numbers is the roadmap, and no wording on this page narrows it.
Cost to a trustworthy referee: 1 u. GR1 the panel. It gates everything else: 2026-08-16 proved the single judge scores green clay above a photoreal-class render, in both directions.
Cost to meet SOTA visual class: 7 u. GR1 panel, GR2 the leap (the splat avatar must beat the ring figure AND the banked WebGL reference on the panel, declared before the experiment), GR3 4D motion, GR8 PBR maps, GR9 pre-integrated SSS -- the visible-surface chain.
Cost to exceed: 13.5 u. Add the Infinigen expansions GR4-GR6 (second species, part-genome, indoor rooms) plus GR7 and GR10-GR13. The exceed axes themselves -- internals, determinism, genome identity, the adult invariant -- are already gate-proven and cost zero new units.
Research bar. Infinigen (Princeton, BSD-3) is measured on procedural nature, creatures, indoor rooms via constraint solve, materials, ground-truth annotation; 2.0 alpha adds controllable APIs and indoor materials and rooms; 331GB flying-indoor video dataset Aug 2026. Theirs: five genome classes, NURBS part lofting, single-surface creatures -- source-level read, grade-1. Ours: our expansion rungs GR4-GR6 and the ground-truth contract flip measured on this page.
Research bar. SIGGRAPH Advances in Real-Time Rendering is measured on twenty years (2006-2026) of the industry's shipped real-time technique. Theirs: the named-technique record. Ours: GR7 mirrors it into the library beside the 9762-paper index; rows then cite talks by name.
Research bar. MetaHuman (Epic) is measured on the external photoreal digital-human bar. Theirs: photoreal artist-authored rigs. Ours: the panel grades our figure against MetaHuman-class stills; our north star, the inside-and-out twin, is a HIGHER bar they do not enter.
42 of 61 capabilities measured|4 of them measured exceeds|19 open|coverage 688/1000|adoption 24 full / 18 partial
Layer 2 · Roadmap
Do this next — computed by the ranker, never chosen by a seat
Order from nx_compare_rank (nx_dr_ocm: (deficit + cost-of-delay + option + enables) x sponsor x self-sufficiency x momentum / cost). FINISH rows are rungs whose symbol is present but whose organ is short of full adoption: the cheapest closures on this board, listed before any new work. Stamp: # asof=1787883535 domain=graphics target_version=0.1 rungs=18 done=5 open=13 finish=1 ranker=nx_dr_ocm
| # | Stage | Rung | Priority | Derivation |
|---|---|---|---|---|
| F | FINISH | Advances archive mirrored (GR7) gp_collect2 | PROMOTED-UNREGISTERED | a real binary nobody can call over MCP: /api/tools/register it; no execution surface runs it either (clock, cron, daemon, roster, actlog and surfaced forks checked) |
| #1 | 0.1 | The leap: genome splat avatar (GR2) gs_avatar_genome | 100 | v=2 m=2 c=40 |
| #2 | later | Pre-integrated SSS (GR9) sr_sss_preint | 2666 | v=4 m=2 c=3 |
| #3 | later | Marschner-class hair shading (GR16) hr_marschner_lobe | 1800 | v=9 m=1 c=5 |
| #4 | later | Linear-light pipeline (GR14) sr_linear_light | 1333 | v=4 m=1 c=3 |
| #5 | later | Skin-bound dynamics (GR13) dyna_bind_skin | 1200 | v=12 m=1 c=10 |
| #6 | later | Second species canon (GR4) bp_species_canon | 700 | v=7 m=1 c=10 |
| #7 | later | Path-traced GI (GR10) gp_bounce_gi | 700 | v=14 m=1 c=20 |
| #8 | later | Native-resolution GPU delivery (GR11) cw_webgpu_compute | 600 | v=6 m=1 c=10 |
| #9 | later | Sovereign strand system (TressFX-referenced) (GR15) hr_strand_sys | 500 | v=10 m=1 c=20 |
| #10 | later | Part-genome creatures (GR5) cg_genome_parts | 350 | v=7 m=1 c=20 |
| #11 | later | Infinigen fork map as data (GR18) if_fork_map | 100 | v=1 m=1 c=10 |
| #12 | later | 4D splat motion (GR3) g4_avatar_rig | 66 | v=1 m=1 c=15 |
| #13 | later | Groom style zones with pinned dynamics (GR17) gr_style_zones | 0 | v=0 m=1 c=15 |
Critical path — contract, done-rule, executor, cost
| Rung | Closes with | Definition of done (pre-declared) | Executor | Est. |
|---|---|---|---|---|
| Referee panel (GR1) | cjc_panel | LANDED 2026-08-19 in one leg (estimate held): cjc_panel composes the statistics judge, the published Kovac skin-gamut ruler, image_beauty and nx_percept, forked per image; FP and FN measured on the labelled corpus v1 (n=10, 3 BAD by operator record, 7 GOOD = the six banked WebGL NXA rig stills captured on the real GPU + the published Mackenzie render) -- stats and symmetry proven non-separating and left report-only, skin and percept separable, floors 85/152 from the margins, composite FP=0 FN=0, ADMITTED under a floor-bound receipt; nx_cjc_panel_gate GREEN, bite-proven both ways. Open inside the rung: the VLM lens is UNOBSERVABLE until a sovereign VLM oracle exists (none on the estate, measured), and the floors need the second-corpus retest (MetaHuman-class stills) before they are believed | Organ | 1 u |
| The leap: genome splat avatar (GR2) after GR1 | gs_avatar_genome | RIGHT-SIZED 2026-08-18 to the published bar (the earlier budget understated by ~10x): v2 by canon-mode bodyfit search (16-dimensional ring-scale under the judge), the gated face pipeline composed in AT >=100K-gaussian face density (GaussianAvatars-class heads), strand-aligned splat GROOM hair at groom scale (85 percent of the industry poly budget), 8K-class skin with wrinkle response; done when the panel grades it above the ring figure AND the banked WebGL reference on the same framing -- declared before the experiment. THE BAR IS NOW DATA (leg 1, 2026-08-19, admitted panel on knowledge/cjc_gr2_framing.list, every row a figcrop): ring figure skin 7 percept -112 BAD; splat v1 skin 118 percept 123 BAD (skin-coloured, perceptually junk-ward); the six WebGL rig crops skin 279-661 percept 210-228 GOOD; Mackenzie skin 624 percept 281. DONE-RULE IN NUMBERS: v2 crop must read composite GOOD with percept above 228 (the best banked rig crop) and skin at or above 279 (the weakest), re-measured with the same conf and receipt | Organ | 4 u |
| 4D splat motion (GR3) after GR2 | g4_avatar_rig | The splat avatar moves with the same gait, breath and expression machinery the mesh uses; replay bit-identical | Organ | 1.5 u |
| Second species canon (GR4) | bp_species_canon | A horse canon through the identical pipeline; a bodybench-style ruler against a CC anatomy oracle of the species proves species-agnostic generation | Organ | 1 u |
| Part-genome creatures (GR5) after GR4 | cg_genome_parts | Typed interpolatable parts over the existing plans; distinct genomes must measure distinct and interpolation must stay anatomically legal | Organ | 2 u |
| Indoor rooms (GR6) | ig_room_solve | Constraint-solved rooms with materials through our engine, deterministic and bounded; REFUSES an unsatisfiable constraint set rather than overlapping furniture | Organ | 2 u |
| Advances archive mirrored (GR7) | gp_collect2 | Second root through the ingest organ, fixed-point crawl, partition reconciles per root; the extractor then cites it | Organ | 0.5 u |
| PBR map set per genome (GR8) | ntb_bake_pbr | Four maps by four regions regenerated per genome at load; normal-map lighting proven through the existing image path | Organ | 1 u |
| Pre-integrated SSS (GR9) after GR8 | sr_sss_preint | Bound to the baked normal map; carries the away-side equality tooth (the T66 form) so no uniform brightener can fake it | Organ | 1 u |
| Path-traced GI (GR10) | gp_bounce_gi | Per-frame bounce beyond cached radiosity; the cache-vs-rebuild equality stays as the determinism tooth | Organ | 2 u |
| Native-resolution GPU delivery (GR11) | cw_webgpu_compute | HALF LANDED 2026-08-18: the WebGL2 tier is LIVE DEFAULT on all five worlds and the WebGPU WGSL world pass is LIVE LEVERED (?gpu=1, real-GPU verified via the CDP wire, boot-stall ceiling hardened); REMAINING: the compute-shader form, the NPC+eye passes to parity so WebGPU can default, and the same frame bit-compared against the CPU path at reference size | Organ | 1 u |
| Expression morphs (GR12) after GR2 | fa_expr_morph | FACS axes over the identity morphs, graded against the 710-expression FACS-class basis MetaHuman ships (the bar, not a smile bone); the panel grades expressions readable | Organ | 2 u |
| Skin-bound dynamics (GR13) | dyna_bind_skin | The proven solver bound to skin vertices, with collision against the body it rides | Organ | 1 u |
Milestones
| Milestone | Rungs | Cumulative | ||
|---|---|---|---|---|
| M0 · Referee trustworthy | GR1 | 1 u | ||
| M1 · The leap published | GR2 | 3 u | ||
| M2 · Living avatar | GR3,GR12,GR13 | 6.5 u | ||
| M3 · Infinigen expanded | GR4,GR5,GR6,GR7 | 11 u | ||
| M4 · Surface and delivery | GR8,GR9,GR10,GR11 | 18 u |
| Rung | Closes with | Definition of done (pre-declared) | Executor | Est. |
|---|---|---|---|---|
| Linear-light pipeline (GR14) | sr_linear_light | All lighting math computed in linear, display-encoded exactly ONCE at output; proven by the tonemap A/B that failed 2026-08-26 (ACES over display-encoded values washed the frame -- the banked double-transfer lesson) now passing: filmic curve applied and the A/B screenshot referee grades it richer, not paler; prerequisite to GR9 | Organ | 1 u |
| Sovereign strand system (TressFX-referenced) (GR15) | hr_strand_sys | The Infinigen-fork hair chain, leg 1: a NishiLang strand system whose sim and render structure is read from the mirrored TressFX source (the canonical shipped open real-time strand system) -- guide strands + per-strand constraints on the one solver, rendered with real width and coverage in our WebGL2 tier; done when a girl-framed capture is graded by the panel above the current ribbon baseline and no ad-hoc constant lacks a cited row | Organ | 2 u |
| Marschner-class hair shading (GR16) after GR15 | hr_marschner_lobe | Hair BRDF per Marschner 2003 through the Karis 2016 / Tafuri 2019 production reductions (mirrored): dual-lobe spec with shifted primary and secondary highlight, azimuthal roughness; constants pinned as cited conf rows before the shader edit, gate compares lobe positions against the published reduction | Organ | 1.5 u |
| Groom style zones with pinned dynamics (GR17) | gr_style_zones | nx_nxa_groom gains the cosmetology-brief zone model (hairline, part, nape, density map) with strand dynamics constants extracted from the mirrored Super-Helices / Discrete Elastic Rods papers and PINNED as oracle rows (closing the conf's own uncited hair_k_q10 correction record); style presets are parameters, every band cited or derived | Organ | 1.5 u |
| Infinigen fork map as data (GR18) | if_fork_map | The fork strategy is a LIVE census, not prose: a conf mapping each Infinigen subsystem (ProcFunc generators, constraint solver, materials, creatures, ground-truth writers, Infinigen-Sim articulation) to its sovereign NishiLang twin, its governing SOTA reference, its gate, and its measured state -- rendered on this board and re-measured at emit; the differentiators (inside-out anatomy, genome identity, bit-exact determinism, the adult invariant) carried as exceed rows | Organ | 1 u |
Layer 3 · Engineering
How this is scored. Every Nishi mark is measured: the generator reads the real organ source on disk and requires the implementing symbol to exist (no self-grading). A watching tag names the organ and symbol contracted to close a gap — the mark flips itself on the next compare beat when that workstream ships, and the
comparewatch- plane row flips with it. The flip is necessary, not sufficient: it proves the symbol exists, never that the capability is good. The bar is the rung's pre-declared done-rule, proven by its gate — a symbol shipped without the behaviour behind it is a defect, and the flip is exactly what makes that defect visible instead of quiet. Competitor marks record documented capability presence — presence, not depth or scale. Adoption is measured too: every measured row carries where its organ stands on the estate's ladder (source → built → promoted → registered → invoked; libraries by importer reach minus validation importers; gates by the execution surfaces that run them). A row is fully adopted only at the top of its ladder; anything short is tagged partial with the exact remedy, so a build nobody promoted can no longer read as shipped. Census stamps: importers asof 1787849099, gate census asof 1787855507 (unix seconds; -1 = census absent).Capability matrix — measured against source
◉ leads / measured exceed● present◐ partial○ absent · click any capability for its evidence
| Capability | Nishi | Unreal Engine 5 | Godot | Infinigen | Houdini |
|---|---|---|---|---|---|
| RENDER | |||||
energy-conserving GGX Cook-Torrance PBR, all integerMeasured:gp_ggx exists in runtime/_hdl_build/nx_gx_pbr_ggx_gate.nx, verified at emit. Microfacet D times F times G over 4 NoV NoL, gate-proven on a metallic roughness sweep -- bit-deterministic where float engines drift. [walter2007] Adoption: GATE:BUILT-UNPROMOTED trial=- — PARTIAL: compiled, never promoted to the serving root: /api/promote it, then roster it. | ● | ◉ | ● | ● | ◉ |
SSAO and depth-of-field in the sovereign rendererMeasured:as_ssao exists in runtime/_hdl_build/nx_anat_sov.nx, verified at emit. as_ssao with an attenuation floor that never crushes a cavity to black, plus as_dof; the same organ carries the SSS wrap, tone variation and micro-relief skin stack (sec27-30 at /world/anatomy_gpu). Adoption: REGISTERED-DARK — PARTIAL: callable, authorised, no MCP invocation on record (a direct fork logs the runner, so this is not proof it never ran); no execution surface runs it either (clock, cron, daemon, roster, actlog and surfaced forks checked). | ● | ◉ | ◉ | ● | ◉ |
pre-integrated SSS bound to baked normalsMeasured:sr_sss_blur exists in runtime/nx_sdfrender.nx, verified at emit. LANDED 2026-08-25 -- Penner's SECOND half. âš THE WATCH SYMBOL WAS DELIBERATELY CHANGED FROM sr_sss_preint TO sr_sss_blur: sr_sss_preint ALREADY EXISTED, already wired and already toothed BEFORE this rung (nx_sdfrender.nx:782, called in sdf_shade_ray, pinned by nx_sdfrender_gate T6), so leaving the watch on it would have flipped this row GREEN on a symbol that does not implement the contract -- the name-keyed completion defect this board has already refused once. The shipped half was the CURVATURE contribution (SDF Laplacian); the half that stops the clay read is per-channel NORMALS, and it was absent. Now each channel is lit by its own blur of DETAIL-toward-GEOMETRIC normal, so fine relief scatters (red loses it, blue keeps it) while the coarse surface stays sharp -- sdf_shade_ray banks the geometric normal BEFORE the relief perturbation and specular, rim, fill, AO and shadow all keep it, which is Penner's own split. sr_sss_preint DELEGATES to sr_sss_preint_n so there is one copy of the arithmetic and every existing caller and tooth is bit-identical by construction; the identity dot(lerp(a,b,t),L) = lerp(dot(a,L),dot(b,L),t) means no per-channel normal is built or renormalised in the hot loop, so no isqrt was added. WEIGHTS DERIVED, NONE DIALLED, and zero conf rows were needed: the profile is the existing SSS_WRAP_R0/G0/B0 already cited and pinned by two gates, the relief width is the baker's own NTB_RELIEF_UM/LAMBDA (its comment declares it a stated physical model, not an artist slider), and blur = W/(W+W_relief) = 179/138/86. nx_sdfrender_sss_gate 11/11 GREEN via /api/gate_run exit_code 0. T6 IS THE LOAD-BEARING TOOTH AND ITS BAR IS NOT A CONSTANT -- it calls the SHIPPED wrap-only path at runtime as the incumbent control (354 vs 741 permil), and a plain blur scores 758, PASSING T5 and FAILING T6, so T5 alone is measurably gameable and that is proven rather than asserted. Bite-proven by manual plant (blur ignoring its channel argument -> 7/11 RED exit 1, dying at T0/T2/T6 with the neg-control collapsing to FALSE-POSITIVE), restore proven byte-identical through the WHOLE BUILD CLOSURE rather than by byte count. Sibling nx_sdfrender_gate rebuilt against the change: 6/6 GREEN, T6 values bit-identical. OPEN AND STATED: the renderer still uses PROCEDURAL mesostructure and does not yet SAMPLE the baked map (reach: baked relief 51 vs procedural tilt 18), so SAMPLING the baked normal map is the follow-on rung; and no trustworthy before/after image exists because nx_sdfrender_gate's write_png is not landing and its T5 gates on determinism only and discards the return -- debt 1787678931, and the A/B was therefore published as pixel counts, not as a picture. [penner2011] Adoption: LIB-WIRED importers=46 nonval=25 — fully adopted (top of its ladder). | ● | ◉ | ● | ● | ◉ |
per-frame path-traced dynamic GIOpen — watchingruntime/nx_gi_path.nx : gp_bounce_gi, re-measured on every compare beat. Ship that symbol and this mark flips itself; the comparewatch- plane row flips with it. Lumen is the bar. Ours today is cached-form-factor radiosity at 20.3x less work with cache-vs-rebuild EQUALITY provable because integer determinism -- a statement float engines cannot make. Contract: per-frame bounce plus ReSTIR-class sampling. [bitterli2020] | ○ | ◉ | ◐ | ● | ◉ |
anti-aliased analytic rasterMeasured:aa_render exists in runtime/nx_aa_raster.nx, verified at emit. The supersampled coverage rasterizer behind fonts, bubbles and UI; nonzero-winding union in one pass. Adoption: LIB-WIRED importers=15 nonval=9 — fully adopted (top of its ladder). | ● | ◉ | ◉ | ● | ◉ |
| MATERIAL | |||||
skin albedo from the genome via the published ITA standardMeasured:si_albedo_of_gene exists in runtime/nx_skin_ita.nx, verified at emit. CIE Lab ITA classes dark to very-light with regional offsets; the world keeps its identity in the outfit while the individual owns her skin. Replaced twelve free sRGB constants. [chardon91] Adoption: LIB-WIRED importers=4 nonval=3 — fully adopted (top of its ladder). | ● | ● | ● | ● | ◉ |
anatomy-derived UV unwrapMeasured:ntx_apply exists in runtime/nx_nxa_texc_lib.nx, verified at emit. Bone-anchored unwrap written into the asset, additive and idempotent, regions-from-anatomy exact with sums checked. Adoption: LIB-WIRED importers=7 nonval=4 — fully adopted (top of its ladder). | ● | ◉ | ◉ | ◉ | ◉ |
deterministic albedo atlas bake and samplingMeasured:ntb_bake exists in runtime/nx_nxa_texbake_lib.nx, verified at emit. Gate 11 of 11; atlas and textured views published at /world/anatomy_gpu; 1024 and 2048 proven. [kkguides] Adoption: LIB-WIRED importers=5 nonval=2 — fully adopted (top of its ladder). | ● | ◉ | ◉ | ◉ | ◉ |
full PBR map set regenerated per genomeMeasured:ntb_bake_pbr exists in runtime/nx_nxa_texbake_lib.nx, verified at emit. Albedo alone reads as clay; the measured floor is 4 maps by 4 regions. Rivals ship fixed files -- ours regenerate per genome at load: ntb_bake_pbr LANDED (nx_nxa_texbake_lib: albedo, spec, gloss, relief-derived normal by 4 regions, skin-IOR F0; verified func-not-comment at admission). Adoption: LIB-WIRED importers=5 nonval=2 — fully adopted (top of its ladder). | ● | ◉ | ◉ | ◉ | ◉ |
| BODY | |||||
procedural body from an anthropometric canon, zero scanned assetsMeasured:bg_parse_canon exists in runtime/_hdl_build/nx_body_gen.nx, verified at emit. The canon is DATA rows (rings, relief, layers); BodyParts3D is oracle only, never product. [bodyparts3d] Up to 8.3M tris with offered-vs-emitted demand reporting, so a capped emit can never again pass as complete. Adoption: REGISTERED-DARK — PARTIAL: callable, authorised, no MCP invocation on record (a direct fork logs the runner, so this is not proof it never ran); no execution surface runs it either (clock, cron, daemon, roster, actlog and surfaced forks checked); only referenced by nx_creatorfit -- a mention, not a run. | ● | ◐ | ○ | ◉ | ◐ |
Catmull-Rom spline lofting through canon ringsMeasured:bg_cmr exists in runtime/_hdl_build/nx_body_gen.nx, verified at emit. C1-smooth organic curvature from the same canon; Infinigen's higher-order NURBS lofted from real-reference profiles is the named remaining smoothness edge. Adoption: REGISTERED-DARK — PARTIAL: callable, authorised, no MCP invocation on record (a direct fork logs the runner, so this is not proof it never ran); no execution surface runs it either (clock, cron, daemon, roster, actlog and surfaced forks checked); only referenced by nx_creatorfit -- a mention, not a run. | ● | ● | ○ | ◉ | ◉ |
inside-out layer stack -- bone, muscle, skin, emergent formMeasured exceed:bg_parse_layers in runtime/_hdl_build/nx_body_gen.nx, verified at emit. EXCEED with the why: read from Infinigen source, their creatures are SINGLE-SURFACE; our fat parameter attenuates muscle relief through real nested layers so lean shows the abs and heavy hides them from the SAME muscle layer, A/B proven. [infinigen23] Honest: inner layers are still coarse envelopes, not named muscles. Adoption: REGISTERED-DARK — PARTIAL: callable, authorised, no MCP invocation on record (a direct fork logs the runner, so this is not proof it never ran); no execution surface runs it either (clock, cron, daemon, roster, actlog and surfaced forks checked); only referenced by nx_creatorfit -- a mention, not a run. | ◉ | ○ | ○ | ○ | ◐ |
canon computed from rules -- sex, build, muscularity, seedMeasured:bp_load_dimorph exists in runtime/_hdl_build/nx_body_proc.nx, verified at emit. The Infinigen pattern (generators plus constraints plus search): dimorphism functions plus seeded noise emit the canon the body factory reads; one rule engine for male, female, anime, individuals. Adoption: LIVE — fully adopted (top of its ladder). | ● | ◐ | ○ | ◉ | ◐ |
face features and digits emitted by ruleMeasured:emitDigitP exists in runtime/_hdl_build/nx_body_proc.nx, verified at emit. Nose, lips, chin, brow, ear, eye, finger and toe emitters. Honest: crude against MetaHuman-class sculpts [metahuman] [metahuman-dna] -- sculpt fidelity is the headline gap, and the fix is canon-mode search under the panel, never hand-modelling. Adoption: LIVE — fully adopted (top of its ladder). | ● | ◐ | ○ | ◉ | ◉ |
a second species through the identical canon engineOpen — watchingruntime/_hdl_build/nx_body_proc.nx : bp_species_canon, re-measured on every compare beat. Ship that symbol and this mark flips itself; the comparewatch- plane row flips with it. Infinigen ships five genome classes; we have proven one species canon. Contract: a horse canon through the same pipeline turns the species-agnostic claim into a measurement. | ○ | ○ | ○ | ◉ | ◉ |
seed-parametric creatures -- quadruped, bird, fish plansMeasured:creaturegen_build_plan exists in runtime/nx_creaturegen.nx, verified at emit. Sovereign Infinigen: the layered-anatomy method generalized to animal body plans with seeded proportions; the skin emerges as the smooth-union envelope plus dermal offset. Adoption: LIB-WIRED importers=9 nonval=3 — fully adopted (top of its ladder). | ● | ○ | ○ | ◉ | ◐ |
interpolatable part-genome across creature typesOpen — watchingruntime/nx_creaturegen.nx : cg_genome_parts, re-measured on every compare beat. Ship that symbol and this mark flips itself; the comparewatch- plane row flips with it. Infinigen's genome joins body, head, leg, foot, tail and horn parts and interpolates BETWEEN types; ours has fixed plans. The expansion contract is typed parts with legal interpolation -- their capability, our layers underneath. | ○ | ○ | ○ | ◉ | ◐ |
| ANATOMY | |||||
eight-system internal anatomy rendered inside and outMeasured exceed:as_load_sites in runtime/_hdl_build/nx_anat_sov.nx, verified at emit. EXCEED with the why: MetaHuman and Infinigen render no internals at all; the bar here is a medical atlas. Peel, explode and stereo proven against the CUDA oracle on the same invariants. Honest: organ shapes vs atlas still open. Adoption: REGISTERED-DARK — PARTIAL: callable, authorised, no MCP invocation on record (a direct fork logs the runner, so this is not proof it never ran); no execution surface runs it either (clock, cron, daemon, roster, actlog and surfaced forks checked). | ◉ | ○ | ○ | ○ | ◐ |
articulated rig with deterministic physics on the real skeletonMeasured:rg_fk_y exists in runtime/_hdl_build/nx_anat_rig.nx, verified at emit. FK joint tree plus damped-pendulum integrator on the 662k-tri BodyParts3D skeleton; replay bit-identical on all frame checksums. Honest gaps: sagittal hinges only, no constrained rigid-body solver. Adoption: SOURCE-ONLY — PARTIAL: source exists, never compiled: /api/build it. | ● | ◉ | ◉ | ● | ◉ |
| MOTION | |||||
real mocap decoded and retargetedMeasured:m_joints exists in runtime/nx_bvhmotion.nx, verified at emit. BVH and VMD decoded sovereignly, retargeted at 30fps; rivals ship authored clip libraries. Adoption: REGISTERED-DARK — PARTIAL: callable, authorised, no MCP invocation on record (a direct fork logs the runner, so this is not proof it never ran); no execution surface runs it either (clock, cron, daemon, roster, actlog and surfaced forks checked). | ● | ◉ | ● | ○ | ● |
general secondary dynamics -- jiggle, recoil, hair, clothMeasured:sd_step exists in runtime/nx_softdyn.nx, verified at emit. Spring-damper with the stability envelope ENFORCED inside the solver so an unstable config cannot be requested; gate 7 of 7 with an anti-vacuity tooth. Adoption: LIB-WIRED importers=7 nonval=5 — fully adopted (top of its ladder). | ● | ◉ | ● | ○ | ◉ |
one damped-oscillator identifier shared by the solver AND by videoMeasured:rd_analyze exists in runtime/nx_ringdown_lib.nx, verified at emit. SHIPPED 2026-08-25. Until today every fidelity claim on /world/beach was PLANT-against-CITED-BAND: nx_dyna_bind_skin rings the XPBD cage down and compares fn/zeta to gamefeel_oracle.conf. Sound for the plant, silent about the render -- correct fn and zeta do not imply correct amplitude, phase to footfall, or mode shape. The estimator was EXTRACTED VERBATIM out of nx_dyna_bind_skin into a lib so our own plant and real footage cannot be measured by two different rulers; the ringdown consumer was rebuilt and nx_behaveprobe read IDENTICAL first_diff_off=-1. It carries a real third state (RD_UNOBS_NO_SAMPLES / NO_CROSSING / NO_PERIOD / NO_DECAY) so a zero frequency can never be confused with an unmeasurable one, and it keeps the incumbent's DECLARED log2 imprecision rather than silently improving it. nx_dynaoracle_gate 18/18 GREEN exit 0 against an ANALYTIC fixture: a mass-spring-damper the gate integrates itself, so w0=sqrt(k/m) and zeta=c/(2 sqrt(km)) are known in closed form and the teeth ask whether the estimator RECOVERS numbers it was never told -- measured zeta 71 permil against a true 79 inside the DERIVED tolerance of 10, period 100 samples against a true 99 inside the DERIVED sampling bound of 11. Both tolerances are derived from declared imprecision, never picked. Bite-proven: a RIGID signal must refuse with a named third state, and a stiffer spring must read as a higher frequency, so a constant-returning stub fails. [physrig] [softsmpl] Adoption: LIB-WIRED importers=3 nonval=2 — fully adopted (top of its ladder). | ● | ○ | ○ | ○ | ○ |
coarse-to-fine optical flow -- large-displacement tissue trackingMeasured:nx_motion_mean_flow_pyr exists in runtime/nx_motion.nx, verified at emit. SHIPPED 2026-08-25 and it closed a MEASURED blindness. nx_motion.nx has carried Lucas-Kanade since 2026-05 with exactly ONE consumer -- its own test (corpus_complete=1) -- i.e. BUILT+UNWIRED, and nx_dynaoracle is its first production consumer. Its own header declares the small-motion assumption, and the first real subject violated it: on a fixture with a 14-pixel oscillation at 2 Hz and 30 fps (peak 5.9 px/frame) it recovered about 1 px of 14 and the consumer correctly reported UNOBSERVABLE. Tuning the fixture down would have been fitting the test to the tool, so the tool was fixed with the standard remedy. The pyramid COMPOSES the incumbent rather than adding a second flow estimator, and the same edit removed a genuine hot-loop allocation: sobel, convolve, temporal-gradient and LK all gained scratch-reusing _into variants with the allocating forms DELEGATING, so every existing caller is unchanged BY CONSTRUCTION -- proven, not asserted, by nx_behaveprobe on a real consumer reading IDENTICAL/STRONG. Measured after: coverage 214 to 1000 permil on the same clip. Adoption: LIB-WIRED importers=3 nonval=1 — fully adopted (top of its ladder). | ● | ◉ | ● | ○ | ◉ |
METRIC SCALE for video measurement -- the declared limit of the video oracleOpen — watchingruntime/nx_dynaoracle.nx : do_metric_scale, re-measured on every compare beat. Ship that symbol and this mark flips itself; the comparewatch- plane row flips with it. THE CONTRACT THAT CLOSES nx_dynaoracle's OWN DECLARED LIMIT, named here rather than left as a caveat in a header. The oracle states before any number that amplitude in MILLIMETRES is UNMEASURABLE from an uncalibrated clip and prints that string instead of a figure, so tissue_amp_walk_mm 32..52 and tissue_amp_run_mm 110..194 remain CITED and unmeasured against our render. Everything the oracle DOES report is dimensionless BY CONSTRUCTION -- fn, zeta, drive frequency, drive-to-response phase lag, and amplitude as a RATIO of a body landmark -- which is exactly why those survive a clip with no calibration, no known camera and no known subject size. The missing input is a metric body frame. Multi-HMR 2 recovers multi-person whole-body mesh from a single image in the ANNY representation, and ANNY is an Apache-2.0 parametric human covering infants to elders with interpretable controls [anny] [multihmr2] -- a permissive licence, unlike SMPL. THE RULE THIS ROW INHERITS: ORACLE ONLY, NEVER PRODUCT, the same rule this board already applies to BodyParts3D, and the same reason the VLM lens on the referee panel abstains rather than acquits -- a learned model may CALIBRATE our instrument and may never ship inside it. Contract: recover a metric body frame per clip, publish the mm-per-pixel it implies WITH the uncertainty that recovery carries, and convert the oracle's ratios into millimetres so the two cited amplitude bands become MEASURABLE rather than merely citable. Until it lands, amp_mm stays a named refusal and every amplitude claim on this board is a ratio. | ○ | ○ | ○ | ○ | ○ |
DAMPING identified from video -- the parameter the field does not fitMeasured exceed:do_crop in runtime/nx_dynaoracle.nx, verified at emit. EXCEED with the why, and the estate declared this gap itself before filling it: gamefeel_oracle.conf's own closing note records that the 2025-26 differentiable-simulation literature (MonoPhysics, PhysCoRe, EgoPhys) recovers Young's modulus, Poisson ratio and yield stress and NEVER damping, never on human soft tissue. PhysRig fits material prototypes and likewise not damping [physrig]; SoftSMPL learns soft-tissue dynamics from 4D scans, not from monocular video [softsmpl]. Damping is the parameter that governs jiggle decay. nx_dynaoracle measures fn, zeta, drive frequency and drive-to-response phase lag from a frame stack through the SAME estimator as the plant, cancelling camera and body motion by differencing a tissue ROI against a RIGID reference ROI. The noise deadband is not a tunable and there is no conf row for it: it is MEASURED per clip from the reference patch's own top and bottom halves, which share camera and body motion exactly, so their differential is measurement noise BY CONSTRUCTION -- a negative control the caller cannot forget to supply and cannot fudge. END-TO-END MEASURED on a synthetic video fixture with analytically known ground truth: coverage 1000 permil (149/149 pairs), drive frequency 500 mHz recovered EXACTLY, damped fn 2142 mHz against a true 1960 (+9.3 percent), zeta 245 permil against a true 200 (+22.5 percent). DECLARED LIMITS, stated before the numbers rather than discovered after: amplitude in MILLIMETRES is UNMEASURABLE from an uncalibrated clip and the organ prints that string rather than a figure, so tissue_amp_walk_mm stays CITED and unmeasured here; closing it needs a metric parametric-mesh recovery (Multi-HMR/Anny class) used as an ORACLE ONLY, never as product, the rule this board already applies to BodyParts3D [multihmr2] [anny]. Integrating velocity cannot recover the INITIAL displacement, so the first lobe carries an unrecoverable DC offset and is skipped before the log decrement -- frequency is timing and survives an offset, damping is amplitude and does not. ffmpeg is used as a pure DECODER and that break-glass is declared in source with the sovereign VP9/H.264 lane named as its replacement. Adoption: RUN-BY:actlog — fully adopted (top of its ladder). | ◉ | ○ | ○ | ○ | ○ |
dynamics bound to skin verticesOpen — watchingruntime/nx_nxa_dyna.nx : dyna_bind_skin, re-measured on every compare beat. Ship that symbol and this mark flips itself; the comparewatch- plane row flips with it. The solver is gate-proven and the DYNA section already writes into the asset; binding to skin geometry is the open rung. | ○ | ◉ | ◉ | ● | ◉ |
hair secondary motion -- spring-bone chain on the ONE solver, measuredMeasured:hd_ringdown exists in runtime/nx_hairdyn_lib.nx, verified at emit. SHIPPED 2026-08-25, gate GREEN 11/11 exit 0 and bite-proven both ways (valid_mutants=1 not_reached=0 killed=1, restore byte-identical, mutant_live=0). Hair is driven as a CHAIN on the same solver as the tissue -- sd_step_axq stepped per segment, K from sb_k_sym, C from sb_c_for_zeta, plant identified by rd_analyze from nx_ringdown_lib. One solver, one derivation pair, one estimator, all COMPOSED: a second spring integrator would have been the duplicate-ruler defect. sd_step_ax now DELEGATES to a Q8-anchor form because a chain needs sub-unit anchors -- on the model-unit entry point every hop floors the parent's motion away -- and every existing caller stays bit-identical by construction. Measured plant fn 4285 mHz, zeta 22 permil; chain propagation proven by phase lag 83 permil at 2 segments rising to 916 at 12. 60 Hz tick DERIVED from SB_MHZ_TICK_SQ, never written. HONEST: requested zeta is an UPPER BOUND in this solver, not an equality -- sd_step's damping term truncates, so below a velocity threshold no damping applies; the author's first published figure of 66 permil came from linear arithmetic and the RINGDOWN said 22, so the gate asserts monotonicity rather than equality and both places were corrected. [springbones] Adoption: LIB-WIRED importers=3 nonval=1 — fully adopted (top of its ladder). | ● | ◉ | ● | ○ | ◉ |
| RENDER | |||||
scalp groom ANATOMY REFEREE -- nose-arbitered facing, density 4800 floor, length band, conformance, face-cone-clear; negative control is the REAL banked defectMeasured:gg_stats exists in runtime/nx_nxa_groom_gate.nx, verified at emit. v3 GREEN 10/10 (2026-08-27). Facing basis is the RIG CONVENTION anterior-negative cross-checked by mesh density (three geometry heuristics gave three answers; the organ REFUSES on disagreement). Surface tooth is NN-to-mesh with the bound derived per asset from median inter-vert spacing -- its first bin-based form was refuted by its own control (real roots read 764 permil interior). Two banked REAL defects are the negative fixtures (a face-side groom: face-cone 540 permil; a chord-blend interior fill: far 613 permil), and the gate refused three intermediate placements pre-install before root placement was rewritten to the PUBLISHED method (2026-08-27): area-weighted barycentric sampling of the scalp TRIANGLES with face-normal launch, per Blender particle hair and the Infinigen creature hair chain -- every root on a face by construction, no synthesis stage, far 0 permil, uniform coverage. The NN bound itself was recalibrated on measured data after refusing a correct groom (triangle interiors are ~circumradius from every vertex): legit max 941 vs interior-fill median 1,397, bound 5x spacing = 1,130. Adoption: GATE:LIVE trial=- — fully adopted (top of its ladder). | ● | ◉ | ● | ● | ◉ |
| MOTION | |||||
hair GEOMETRY on the cast -- the renderer rung both hair legs needMeasured:gpe_hair_geometry exists in runtime/nx_game_page_emit.nx, verified at emit. THE ASSET WAS NEVER THE BLOCKER, measured 2026-08-25. All four corpus FBX carry groomed hair (dark_witch 22386v/38016t, paladin 19252v/29914t, dark_knight 16583v/28416t, toon3d8 309913v) and nx_nxa_hair extracts them correctly -- it was built for the FIRST TIME today and a null-deref on its first ever run was fixed. The blocker is the CONSUMER: HVRT appears exactly 3 times in 23,221 sources (corpus_complete=1) and ALL THREE are inside nx_nxa_hair.nx itself, so NOTHING READS IT. Every site page is emitted by an organ, so no page can render hair geometry; on /world/beach hair is still a GLSL colour override at nx_game_page_emit.nx:480 over a crown-and-back-of-skull band. The dynamics leg lands on the same wall: nx_hairdyn_lib is imported only by its gate, sb_tick_fn still drives SB_HAIR as ONE point, and the renderer consumes ONE offset where a chain emits N. ★★★★★★HAIR IS ONE DEFECT WEARING TWO MASKS -- an ingest with no consumer and a chain solver with no renderer -- AND BOTH LEGS CONVERGE ON A RENDERER RUNG, NOT A PHYSICS ONE. Adoption: REGISTERED-DARK — PARTIAL: callable, authorised, no MCP invocation on record (a direct fork logs the runner, so this is not proof it never ran); no execution surface runs it either (clock, cron, daemon, roster, actlog and surfaced forks checked). | ● | ◉ | ● | ○ | ◉ |
FACS facial expression morphsMeasured:fa_expr_morph exists in runtime/nx_faceanat.nx, verified at emit. Identity morphs exist as anthropometric axes; expression axes do not. Adoption: LIB-WIRED importers=12 nonval=2 — fully adopted (top of its ladder). | ● | ◉ | ● | ○ | ◉ |
| SPLAT | |||||
sovereign integer 3D gaussian splatting, anisotropic, depth-sortedMeasured:gs_render_aniso exists in runtime/nx_gsplat.nx, verified at emit. The frontier photoreal representation (GaussianAvatars, HumanSplat lineage); all integer with an exp-LUT, front-to-back compositing. [kerbl3dgs] Adoption: LIB-WIRED importers=19 nonval=6 — fully adopted (top of its ladder). | ● | ◐ | ○ | ○ | ◐ |
differentiable color fit step toward trained splatsMeasured:gs_color_grad_step exists in runtime/nx_gsplat.nx, verified at emit. 3DGS at SOTA is a TRAINED fit, not just a splat renderer; the gradient step is the first tooth of that loop. Adoption: LIB-WIRED importers=19 nonval=6 — fully adopted (top of its ladder). | ● | ◐ | ○ | ○ | ◐ |
hybrid -- outside splat, inside parametric meshMeasured:gb_bind exists in runtime/_hdl_build/nx_gsplatbind.nx, verified at emit. The mid-2026 industry architecture (SplattingAvatar class), operator-ordered 2026-07-27: splats bound barycentrically to triangles, the mesh deforms and the splats follow. [splattingavatar24] Built; the leap organ composes it. Adoption: SOURCE-ONLY — PARTIAL: source exists, never compiled: /api/build it. | ● | ○ | ○ | ○ | ○ |
genome-driven splat avatar -- THE LEAPOpen — watchingruntime/nx_gsplat.nx : gs_avatar_genome, re-measured on every compare beat. Ship that symbol and this mark flips itself; the comparewatch- plane row flips with it. Operator-resequenced to the front. v1 ran end to end (94,416 tris to 87,036 gaussians to portrait) and the operator's verdict on v1 stands recorded; the work relocated to canon-mode search under the judge, composing the gated face pipeline, and hair volume (85 percent of the industry reference's polygon budget is hair). Same watch as the charsim board -- one contract, two pages flip. THE BAR BECAME DATA 2026-08-19 (GR2 leg 1): under the admitted referee panel on one figure-crop framing (nx_charjudge_census figcrop, list knowledge/cjc_gr2_framing.list) the ring figure reads skin 7 / percept -112 BAD, the v1 splat skin 118 / percept 123 BAD, the six banked WebGL rig crops skin 279-661 / percept 210-228 GOOD, Mackenzie 624 / 281 -- so v2 is done when its crop reads GOOD with percept above 228 and skin at or above 279 on the same conf and receipt. [gaussianavatars24] [infinihuman25] | ○ | ○ | ○ | ○ | ○ |
4D animated splats driven by the rigOpen — watchingruntime/nx_gsplat4d_lib.nx : g4_avatar_rig, re-measured on every compare beat. Ship that symbol and this mark flips itself; the comparewatch- plane row flips with it. One animation system, two render fidelities -- an architecture MetaHuman cannot have because its animation is welded to its mesh. | ○ | ○ | ○ | ○ | ○ |
| WORLD | |||||
five live wasm worlds, sovereign-only renderMeasured:wc_mob_mesh exists in runtime/nx_wasm_craft.nx, verified at emit. /world/craft, shore, vale, isle, crags: NishiLang compiled to wasm, CPU raycaster, no GPU shader in the path; skinned genome-distinct characters in-world, gate 71 of 71 with eyes, wrap-SSS and the hourglass rework all toothed. Adoption: LIB-WIRED importers=7 nonval=3 — fully adopted (top of its ladder). | ● | ◉ | ◉ | ○ | ○ |
procedural worldgen with real lighting and hydrologyMeasured:wp_shade exists in runtime/nx_worldpipe.nx, verified at emit. Sun-facing diffuse off the real heightfield, gate-proven in both directions; hydrology traces downhill to the sea. Honest: coherent but thin against Minecraft-class staging. Adoption: LIB-WIRED importers=8 nonval=3 — fully adopted (top of its ladder). | ● | ● | ● | ◉ | ◉ |
indoor rooms by constraint solve -- the Infinigen 2.0 expansionMeasured:ig_room_solve exists in runtime/nx_indoorgen_lib.nx, verified at emit. SHIPPED 2026-08-17/18, their 2.0 alpha frontier through OUR engine: nx_indoorgen composes the estate part-solver (envelope, noclip, door-clearance, wall-hug rows), quantizes to voxel cells under a pre-declared accept rule and REFUSES an unsatisfiable set (neg-control toothed) rather than overlap furniture; SITING is solved too (flattest-anchor search per site cell + foundation skirt -- the rolled-anchor stamp it replaced rejected every vale cell, measured before fixed), and the engine stamps the solved layouts -- cottages on craft, driftwood huts on the shore, a 4-building stone-cot village in the vale -- proven by nx_indoorgen_gate 8/8 and craft-gate 74/74 (T72 structure, T73 seam determinism, T74 density-as-data). SHIPPED LIVE 2026-08-18: reconciliation done (the 21 AHEAD strings were superseded rewords of nx_parse diagnostics, proven by fixture), nx_compile_wat promoted at 364 pages, and nx_craft_emit shipped all 5 recipe worlds (craft/beach/vale/isle/crags) vm-gate 5/5, page_verify GREEN; the module-owned page-sizing MECHANISM is shipped in both backends (wat_mem_pages reads a declared static, guards bite-proven 0/70000/5) but the ENGINE declaration is blocked by an nxasm gap (nonzero-init writable statics unsupported both lanes -- debt filed with the unblock spec); until it lands the vm-gate FITS tooth is the drift catcher, and it is the tooth that caught the original 192 drift. [infinigen-indoors24] [infinigen-flying26] [infinigen-docs] Adoption: LIB-WIRED importers=2 nonval=1 — fully adopted (top of its ladder). | ● | ● | ● | ◉ | ◉ |
ground-truth dataset emission -- RGB, depth, normals, semanticsOpen — watchingruntime/nx_world_snap.nx : ws_groundtruth_emit, re-measured on every compare beat. Ship that symbol and this mark flips itself; the comparewatch- plane row flips with it. Infinigen's annotation pipeline is the reason its datasets exist; our renderer already holds depth and normals per pixel. Contract: an emit verb publishing aligned channels, making the estate a dataset factory. THE RIVAL BAR BECAME CONCRETE 2026-08-25, read from NAVER LABS Europe's rendered code index (the raw fetch carried it; an earlier claim that this page was a JS shell was MY OWN keyword-filtered grep failing, not the fetch): Syn4D is a multiview synthetic 4D dataset of dynamic scenes covering camera pose, depth, dynamic 3D reconstruction, 2D and 3D tracking, human pose and novel-view synthesis [syn4d], and Anny-One is a large-scale synthetic dataset for 3D human understanding with multi-person configurations, multi-view setups at extreme camera positions, diverse FOVs, indoor scenes and full ANNY annotations [anny]. PhysRig reached the same wall from the modelling side and had to BUILD a synthetic corpus because, in its own words, evaluating physics-based skinning is hard for lack of suitable benchmark datasets [physrig]. THE EXCEED IS STRUCTURAL, NOT A BIGGER NUMBER: every one of those ships a FIXED corpus, while ours would emit from the genome and the recipe planes -- parametrically unbounded, seed-addressable and regenerable rather than downloaded, with the ground truth exact because the generator knows it rather than estimated because a labeller guessed it. That is the Infinigen-2 procgen integration the operator asked for, arriving AFTER the output is nailed down and not before: the measure-then-generate order, with nx_dynaoracle and the referee panel as the rulers that decide when the output IS nailed down. THE 2026 BAR IS NOW CITED AND PINNED: Syn4D [syn4d] is a multiview synthetic 4D dataset whose procedural multi-camera pipeline jointly emits RGB, depth, instance segmentation, optical flow, dense long-range 3D tracks and parametric SMPL-X pose, all spatially and temporally aligned, so any pixel unprojects to 3D at any frame and reprojects into any other camera. That cross-view consistency is the part our single-camera snap does not yet claim, and it is the honest distance this contract has to close rather than a box we can tick by emitting channels alone. | ○ | ● | ○ | ◉ | ◐ |
native-resolution delivery on the visitor's GPUOpen — watchingruntime/nx_craft_emit.nx : cw_webgpu_compute, re-measured on every compare beat. Ship that symbol and this mark flips itself; the comparewatch- plane row flips with it. A CPU raycaster cannot reach 1200x750 in 24ms on typical hardware (measured in the 2026-08-17 delivery fixes); contract: OUR raycaster emitted as a compute shader -- the visitor's GPU executes our code, sovereignty kept. | ○ | ◉ | ◉ | ○ | ○ |
| REFEREE | |||||
honest body ruler with an anti-Goodhart detail judgeMeasured exceed:bb_detail_score in runtime/_hdl_build/nx_bodybench.nx, verified at emit. EXCEED with the why: headline = MIN of front, side, shape and DETAIL; the detail judge is IoU-in-variance-space, non-gameable by noise (the FBM Goodhart scores 0) and subdivision-proof (1.53M tris moved it nowhere). No rival ships an equivalent honesty instrument. Adoption: REGISTERED-DARK — PARTIAL: callable, authorised, no MCP invocation on record (a direct fork logs the runner, so this is not proof it never ran); no execution surface runs it either (clock, cron, daemon, roster, actlog and surfaced forks checked); only referenced by nx_bodydensity_gate -- a mention, not a run. | ◉ | ○ | ○ | ◐ | ○ |
mechanical character-image judge with full-population censusMeasured:cjc_hw exists in runtime/_hdl_build/nx_charjudge.nx, verified at emit. corpus=169 mean=138, the number every beauty claim must beat. Proven blind in BOTH directions (clay 257, photoreal-class 0) -- which is exactly why the panel contract below gates all fidelity claims. Adoption: LIVE — fully adopted (top of its ladder). | ● | ○ | ○ | ○ | ○ |
multi-lens judge panel that can see beautyMeasured:cjc_panel exists in runtime/nx_charjudge_census.nx, verified at emit. Phase 0 of past-MetaHuman. LANDED 2026-08-19 (GR1, the ranker's #1): cjc_panel in nx_charjudge_census composes the incumbents -- nx_charjudge (stats), nx_color_kovac_skin_pixel (published skin gamut over the centre box), nx_image_beauty path (symmetry), nx_percept judge (relative perceptual position vs the 10 human-confirmed commercial frames) -- forked per image, numbers always, verdicts ONLY under an admission receipt that calibrate writes at FP=0 FN=0 and binds to the exact floors (edit a floor = unadmitted by construction). CALIBRATION RUN 1 on the labelled corpus v1 (n=10: 3 BAD = the operator's green gumby crop, the recorded v1 splat verdict, a figureless world snapshot; 7 GOOD = 6 banked WebGL NXA rig captures from /world/rigview on the real GPU + the published Mackenzie render): stats good_min 70 vs bad_max 234 NOT separable (the blind ruler, measured again), sym 650 vs 1000 NOT separable -- both report-only; skin 130 vs 40 and percept 175 vs 129 SEPARABLE -- floors 85 / 152 from the margins, composite FP=0 FN=0, ADMITTED. nx_cjc_panel_gate GREEN against the deployed binary and bite-proven both ways (the green twin fails the skin lens; a moved floor un-admits; a mislabelled corpus refuses and revokes the receipt), one ruler in runtime/_hdl_build (a same-day sibling twin in runtime/ was retired reversibly). HONEST HORIZON: the VLM lens is UNOBSERVABLE -- no sovereign VLM oracle exists on the estate (tree grep + local probe 2026-08-19), the slot abstains and never acquits; n=10 is a first corpus and the second-corpus retest is owed before the floors are believed; the skin lens is framing-sensitive (central box) so verdicts compare like framings, which the GR2 done-rule already requires. SECOND-CORPUS RETEST 2026-08-19 (corpus v2 = knowledge/cjc_panel_labels.list: BAD gumby crop + v1 splat, GOOD = two PUBLISHED MetaHuman hero stills fetched sovereignly and banked, the far bar the plan names): stats 103 vs 234 and sym 488 vs 1000 again NOT separable; percept 212 vs 126 SEPARABLE with floor 152 inside the band (fp=0 fn=0, consistent with v1); but the Kovac DAYLIGHT skin rule reads the cinematically graded MetaHuman skin at 0-6 permil (centre box) and 0-4 (figure mask), so floor_skin=85 fails both GOOD refs (fn=2) and the retest was RECORDED as having REFUSED and revoked the receipt. -- CORRECTED 2026-08-25, AND BOTH HEADLINE CLAIMS OF THAT RETEST WERE WRONG. (1) IT NEVER REVOKED ANYTHING: it ran in `panel` mode and ONLY `calibrate` writes a receipt, so knowledge/status/cjc_panel_admission.txt still read admitted=1 at floors none,85,none,152 six days later and the live panel printed admission=ADMITTED -- a referee actively calling BOTH published MetaHuman references BAD while this board recorded it as safely revoked. The panel's own law that editing a floor un-admits it fires on a floor EDIT; a corpus that REFUTES the floors revokes nothing. ★★★★★★A REVOCATION NOBODY PERFORMED DID NOT HAPPEN, AND A NOTE ASSERTING A SAFETY STATE THE ARTIFACT DOES NOT CARRY IS WORSE THAN NO NOTE -- it is the stale-RED-gate class inverted, a stale GREEN nobody re-measured. (2) GRADING WAS NOT THE MECHANISM AND THE NAMED REMEDY WAS BUILT AND REFUTED: two published chroma-only rules with no luminance term (Kovac Cb/Cr with Y dropped; Chai and Ngan 1999) were measured over the FULL 3x3 rule-by-region grid on the union -- NINE OF NINE NOT SEPARABLE, good_min=0 in every cell, and dropping the luminance gate never raised good_min, it only inflated bad_max (centre 52 to 848/903). The cause is ONE CORPUS ROW: cmp_graphics_metahuman_ref0_rgb.png reads exactly 0 skin in all nine cells and nx_charjudge's own face axis reads cjp_face=0 -- no face in a GOOD row of a CHARACTER corpus. Its banked original measures identically, so the conversion is faithful, and cjc_ref_mackenzie.png (also an external published render) scores 624, so 'published external reference' is not the discriminator. RE-ADMITTED 2026-08-25 on the UNION v1 u v2 (knowledge/cjc_panel_labels_union.list, n=13 = 9 GOOD / 4 BAD, partition reconciled): the skin floor is REMOVED and skin is REPORT-ONLY, percept is the sole voting lens at floor 152 (173 vs 129, margin 44, fp=0 fn=0), and the receipt is rewritten BY THE CALIBRATE PATH now binding skin_lens and skin_rule as well as the floors, so changing region or ruler un-admits by construction. nx_cjc_panel_gate 13/13 GREEN exit 0 against the DEPLOYED binary and bite-proven both ways (T4 wrong floors refuse / matching admit with the green twin failing the skin lens; T5 swapped labels REFUSE; T10 a refused recalibration rewrites the receipt to admitted=0; new T11 matching floors on another region or ruler do NOT admit; new T12 an unrecognised skin_rule REFUSES by name). HONEST LIMITS: percept is now the ONLY voting lens -- a single-lens referee until a second earns a floor, a real reduction in strength stated rather than hidden; the VLM lens stays UNOBSERVABLE and abstains, never acquits; n=13 measures separability, not generalisation; and excluding the two MetaHuman rows the skin lens WOULD separate at region=fig (GOOD 373-833 vs BAD 6-286), so a second voting lens is recoverable if the corpus owner replaces ref0 -- the rows were deliberately NOT relabelled or dropped, because dropping the rows that refute a floor is exactly the relabelling this corpus forbids. [metahuman-ref0] [metahuman-ref1] Adoption: LIVE — fully adopted (top of its ladder). | ● | ○ | ○ | ○ | ○ |
head-to-head vs-ruler -- parity is not betterMeasured:vs_verdict exists in runtime/_hdl_build/nx_vsbench.nx, verified at emit. beat_index 0 permil over 10 cross-class cells. This page therefore claims breadth and mechanisms, never perceptual wins. Adoption: REGISTERED-DARK — PARTIAL: callable, authorised, no MCP invocation on record (a direct fork logs the runner, so this is not proof it never ran); no execution surface runs it either (clock, cron, daemon, roster, actlog and surfaced forks checked). | ● | ○ | ○ | ○ | ○ |
| RESEARCH | |||||
9762-paper citable graphics index, mirrored non-rottablyMeasured:gx_scan exists in runtime/nx_gfxpapers_extract.nx, verified at emit. SIGGRAPH, EG, I3D, EGSR, SCA, SGP, HPG, PG 2000-2026 from the Ke-Sen mirror in knowledge/library (partition anchors 9984 reconciles); any row here can cite a named paper instead of an adjective. [kesen-mirror] Adoption: PROMOTED-UNREGISTERED — PARTIAL: a real binary nobody can call over MCP: /api/tools/register it; no execution surface runs it either (clock, cron, daemon, roster, actlog and surfaced forks checked). | ● | ○ | ○ | ○ | ○ |
SIGGRAPH Advances course archive mirrored beside itMeasured:gp_collect2 exists in runtime/nx_gfxpapers_ingest.nx, verified at emit. LANDED 2026-08-17: gp_collect2 plus gp_root_of turned the ingest into a second-root mirror -- anchor-driven collection, rebuild host derived from the index url, subdir-safe names; the Advances crawl banked 25 of 26 reachable pages, partition RECONCILED, the one dead link already rotted at capture. Open tail: the extractor citing adv_ pages into the paper index. [advances-archive] [karis2021] Adoption: PROMOTED-UNREGISTERED — PARTIAL: a real binary nobody can call over MCP: /api/tools/register it; no execution surface runs it either (clock, cron, daemon, roster, actlog and surfaced forks checked). | ● | ○ | ○ | ○ | ○ |
| DELIVERY | |||||
GPU last mile in third-party browsers (WebGL2 tier LIVE)Measured:gpe_build exists in runtime/nx_game_page_emit.nx, verified at emit. LIVE 2026-08-18 BY OPERATOR ORDER: GPE_EMIT_WEBGL flipped on -- every /world page now renders on the visitor's GPU at device resolution (voxel world uploaded as a 3D R8UI texture, shader passes for world plus skinned characters) with the sovereign NishiLang-wasm rasteriser ALWAYS compiled in as automatic fallback and ?gl=0 lever. The 08-16 sovereign-only order still binds the charsim QUALITY ladder: quality is measured on the sovereign path; the GPU is the delivery cable. Godot leads on first-class web export; UE5 web export is community-tier. [webgpu-status26] Adoption: REGISTERED-DARK — PARTIAL: callable, authorised, no MCP invocation on record (a direct fork logs the runner, so this is not proof it never ran); no execution surface runs it either (clock, cron, daemon, roster, actlog and surfaced forks checked). | ● | ● | ◉ | ○ | ○ |
WebGPU tier -- WGSL world pass over the same organ stateMeasured:gpe_emit_webgpu exists in runtime/nx_game_page_emit.nx, verified at emit. STAGE A LANDED 2026-08-18, LEVERED ?gpu=1: the organ emits the full WGSL world pass (line-faithful FSH transliteration incl. the fragment-origin Y-flip) plus WebGPU init/draw with a BOOT-STALL CEILING -- both headless Chromiums on the dev box measured wedging INSIDE requestAdapter/createShaderModule with no rejection, so the race caps a visitor's cost at GPU_BOOT_MS and the ladder falls through to GL/CPU. Verified on REAL GPU via the CDP wire: page doors gpu:1/fail:null/px-dG, frame judged RENDERED (ink 870 permil, 103 distinct colors), pixel-faithful to the GL tier. DEFAULT stays GL until the NPC and eye passes reach parity (stage B) -- characters are GL programs today, and a default that loses the girls would be a regression wearing a newer API. WebGPU ~82 percent global; no RT extension in spec (web RT stays compute-shader); subgroups behind flags, bindless in proposal. [webgpu-status26] [webgpu-roadmap26] Adoption: REGISTERED-DARK — PARTIAL: callable, authorised, no MCP invocation on record (a direct fork logs the runner, so this is not proof it never ran); no execution surface runs it either (clock, cron, daemon, roster, actlog and surfaced forks checked). | ● | ● | ● | ○ | ○ |
shaders compiled from NishiLang (nx_wgsl backend)Measured:wgsl_emit_module exists in runtime/nx_wgsl.nx, verified at emit. The sovereign move: our compiler emits WGSL exactly as it emits WAT (sibling backend of nx_wasm.nx), so GPU shaders are written in NishiLang and one source of truth drives the CPU rasteriser, the WebGL/WebGPU tiers, and later the native NishiOS pipe. Rivals compile HLSL/GLSL through vendor toolchains (tint/naga); nobody compiles their own language end-to-end. [wasm3-state26] Rung 2b recovers STRUCTURED CONTROL FLOW from the CFG -- if-diamonds, while loops, loop-containing-if and nested loops -- with a recursive region reducer, on the MEASURED precondition that every loop header in the shader corpus is single-entry, refusing multi-entry and irreducible graphs by name rather than mis-shaping them [relooper-zakai11] [llvm-cfgstackify] Adoption: REGISTERED-UNAUTHORISED — PARTIAL: registered, no cap ever minted; no execution surface runs it either (clock, cron, daemon, roster, actlog and surfaced forks checked); only referenced by nx_wgsl_gate -- a mention, not a run. | ● | ● | ● | ○ | ○ |
WORLD-ROADMAP (BEAUTY TIER): film-quality environments -- the actual Houdini-exceed laneOpen — watchingruntime/nx_gsplat.nx : env_beauty_scene, re-measured on every compare beat. Ship that symbol and this mark flips itself; the comparewatch- plane row flips with it. EH0, THE TIER CORRECTION (operator 2026-08-18: voxel was supposed to be ONE type, not the ceiling): the voxel worlds are the GAME tier and the rungs below improve them, but Houdini-exceed is a BEAUTY tier from the same genome/recipe planes -- full-detail SDF/mesh terrain at Infinigen scale (the estate already meshes 8.3M tris in body_gen and walks SDF-to-mesh in the skull lane; the missing piece is the TERRAIN-scale SDF world lane), rendered as gaussian-splat scenes (nx_gsplat, gate-proven) under the path-GI contract above, with ground-truth emission making it a dataset factory. THE 2026 CEILING IS LEARNED: SIGGRAPH 2026's top-attention terrain work is a diffusion generator with procedural-noise properties (seed-consistent, constant-time random access, infinite) -- so the ladder's top rung is a LEARNED terrain generator on the estate's integer nofloat ML stack, graded against that class, with classical erosion+climate below it as the sim substrate the learned tier trains and grades against. One engine, many representations: voxel game tier, heightfield live tier, SDF/mesh+splat beauty tier, all one seed. [terraindiff26] [infinigen23] | ○ | ● | ● | ◉ | ◉ |
WORLD-ROADMAP (GAME TIER): hydraulic+thermal erosion on the heightfieldMeasured:wp_erode exists in runtime/nx_worldpipe.nx, verified at emit. EH1, THE LARGEST READS-AS-REAL JUMP (gap ladder 2026-08-18): value-noise terrain reads as blobs; carved valleys, ridgelines, talus and drainage basins are what reads as terrain. Pipe-model shallow-water hydraulic erosion plus thermal talus on the worldpipe heightfield -- integer-exact, which turns MASS CONSERVATION into a provable gate tooth instead of a float hope, and judged against PUBLISHED geomorphology bands (drainage density, hypsometric curve) exactly as T53 judges trees per hectare. Houdini's signature terrain feature. CORRECTION 2026-08-24: the earlier note here read 'catalogue-proven ABSENT (capsearch: no erosion organ)' -- that was a capsearch ZERO read as absence, and capsearch indexes registered tools only, so it is structurally blind to every library. The full Mei 2007 pipe model (all seven stages, four regimes, Q14) EXISTS in runtime/nx_water_erosion.nx and has since 2026-05; it was SOURCE-ONLY and is now gated on /compare/procgen with mass conservation as a tooth. CLOSED 2026-08-25 (PG20): the INTEGRATION landed. wp_erode in runtime/nx_worldpipe.nx bakes the terrain into the grid the hydrology stage already uses, drives the full Mei pipeline plus a generalised Musgrave talus pass through ONE owner (runtime/nx_terrain_erode_lib.nx), and stores a delta the height function composes. Proven by nx_worldpipe_erode_gate 21/21 GREEN, exit_code 0, BITE-PROVEN, over the full 102400-cell population: mass exact, INERT before the bake so no existing consumer moved, 101614 cells changed after it, causality restoring 102400 of 102400 with the stage off, and two bakes byte-agreeing everywhere. EH1's own phrase 'judged against PUBLISHED geomorphology bands' is HALF DELIVERED and the half is named: Strahler's hypsometric integral is the referee and is proven non-gameable because it is dimensionless (offset-invariant EXACTLY, scale-invariant within a derived truncation bound), reading 356 to 472 permille; DRAINAGE DENSITY is refused as a subject mismatch because km per km2 needs a metric scale this world does not have, and no maturity band edge is quoted because the primary answered HTTP 403 Cf-Mitigated challenge. [mei2007] [musgrave1989] Adoption: LIB-WIRED importers=8 nonval=3 — fully adopted (top of its ladder). | ● | ● | ● | ◉ | ◉ |
WORLD-ROADMAP (GAME TIER): climate to biome to competition fieldsOpen — watchingruntime/nx_worldpipe.nx : wp_biome_field, re-measured on every compare beat. Ship that symbol and this mark flips itself; the comparewatch- plane row flips with it. EH2: today's scatter is hash-per-cell with genome species; the Houdini/Infinigen form derives temperature and moisture from elevation, latitude and the rain the calendar ALREADY simulates, then species compete for light, water and slope. Most inputs exist in the engine (biome banding, snow line, seasons, rain); the field layer is the gap. Turns trees-placed into an ecosystem grown, all recipe-plane data. [infinigen23] | ○ | ● | ● | ◉ | ◉ |
WORLD-ROADMAP (GAME TIER): volumetric clouds and fog on the GPU tierOpen — watchingruntime/nx_game_page_emit.nx : gpe_cloud_volume, re-measured on every compare beat. Ship that symbol and this mark flips itself; the comparewatch- plane row flips with it. EH3: clouds are 2D hash sheets in the shader today. Real 3D noise-field volumetrics raymarched at device resolution became FEASIBLE THIS SESSION -- the WebGL2/WebGPU tiers are exactly where they run; the sovereign CPU tier keeps its sheet clouds as the floor. [webgpu-status26] | ○ | ● | ● | ● | ◉ |
WORLD-ROADMAP (GAME TIER): road and path networks between settlementsOpen — watchingruntime/nx_wasm_craft.nx : wc_road_net, re-measured on every compare beat. Ship that symbol and this mark flips itself; the comparewatch- plane row flips with it. EH4: settlements landed 2026-08-18 (constraint-solved siting, 74/74); roads are the missing connective tissue -- the same part-solver lane that sites buildings routes paths (slope-cost shortest paths between anchors, stamped as data). Ladder order EH1-EH4 declared: erosion first, then biome fields feeding the existing scatter, then volumetrics, then roads; GI (gp_bounce_gi), dataset factory (ws_groundtruth_emit) and GPU compute (cw_webgpu_compute) rows above already carry the render half of the same roadmap. | ○ | ● | ● | ● | ◉ |
| NATIVE | |||||
NishiOS GPU raster rung on the proven sovereign compute laneOpen — watchingruntime/_hdl_build/_gpu_dxg_submit_gate.nx : gpu_raster_rung, re-measured on every compare beat. Ship that symbol and this mark flips itself; the comparewatch- plane row flips with it. From the first bit up, measured not asserted: the dxg ladder R1-R6 already resolves adapter-to-context-to-resident-VA and has a REAL submission ACCEPTED by the RTX 5080 (ABI measured by recon sweep, tamper matrix, repeatable), and the bare-metal lane carries PCI/GMMU/GPFIFO/doorbell/QMD-v5 gates with golden compute and matmul binaries. The raster contract: a compute-shader rasteriser writing a framebuffer through that same lane -- the native twin of the browser tiers, no libvulkan/libd3d12 anywhere. [webgpu-status26] | ○ | ◉ | ◉ | ● | ◉ |
FOUNDATION (meta, no rival claim): first-byte-up chain for every rung on this board -- what it rests on, where it stops being sound, and which unsound link you can actually act onMeasured:fb_rank_class exists in runtime/nx_firstbyte.nx, verified at emit. META ROW: this is the board's own soundness check, not a rendering capability, so it claims no exceed and scores no rival -- a dependency analyser measured against Houdini would be a subject mismatch. Joins every row here to the estate's own import graph (18776 nodes, 51163 edges, via eg_ancestors) and asks the cap census per module by EXIT CODE (fb_cap_query; 69 forks deduped across 141 link queries), so it derives no edges of its own and adds no second ruler. Measured 2026-08-23: 48 rows partitioning to ready 0, capped 27, placeholder 21, absent 0, empty-chain 0, unknown 0, with the sum printed and checked; 114 chain links, deepest 13. ALL 124 capped links are reported per rung and RANKED, never just the nearest -- own 23, rare 47, ubiquitous 54 -- and both boundaries are DERIVED rather than picked: OWN is identity with the rung's own source, UBIQUITOUS is presence in every resolved chain on this board. A cap in nx_syscalls (fan-in 15042) is an estate-wide concern and never a rung's blocker; a cap in the rung's own module belongs to the lane that owns the rung. ready=0 is a TRUE FINDING, not a broken instrument: only 2.56 percent of estate capacity bounds are DERIVED today, so almost no full transitive chain is clean. Proven by nx_firstbyte_gate 28/28 GREEN -- fbg_rank_order_ok parses all 27 ranked lists and 124 entries and asserts class order with ascending fan-in inside a class, plus a bite-proven neg-control that an absent matrix REFUSES by name instead of reporting an empty healthy board. Validation against the five known foundation defects of 2026-08-23: nx_gsplat CAUGHT and attributed OWN, the nx_skeleton mask LIFTED after the rig-floor lane derived that bound away (its census now reads capacity_population=0 verdict=CLEAN, and a link the census calls clean is never named as unsound), and three declared MISSES that an import graph structurally cannot see -- a runtime DATA prior, a language-level gap, and a capability no row declared. Data-dependency edges are the declared next axis and are printed as NOT-DERIVED on every run. Adoption: REGISTERED-UNAUTHORISED — PARTIAL: registered, no cap ever minted; no execution surface runs it either (clock, cron, daemon, roster, actlog and surfaced forks checked); only referenced by nx_firstbyte_gate -- a mention, not a run. | ● | ○ | ○ | ○ | ○ |
| RENDER | |||||
linear-light pipeline, display-encoded once at outputOpen — watchingruntime/nx_game_page_emit.nx : sr_linear_light, re-measured on every compare beat. Ship that symbol and this mark flips itself; the comparewatch- plane row flips with it. Every rival engine computes lighting in linear and tonemaps once; our page pipeline is display-encoded throughout, measured 2026-08-26 when an ACES pass double-applied the transfer and washed the frame. Prerequisite to the SSS rung. | ○ | ● | ● | ● | ● |
pre-integrated subsurface skinOpen — watchingruntime/nx_game_page_emit.nx : sr_sss_preint, re-measured on every compare beat. Ship that symbol and this mark flips itself; the comparewatch- plane row flips with it. Penner pre-integrated scattering LUT baked integer-side like the sun LUT, bound to the baked normal map; the away-side equality tooth so no uniform brightener can fake it. Was rung GR9 with no matrix row -- the ranker read it as value 0; this row is its value carrier. | ○ | ● | ● | ● | ● |
| HAIR | |||||
sovereign strand system vs TressFX barOpen — watchingruntime/nx_softbind.nx : hr_strand_sys, re-measured on every compare beat. Ship that symbol and this mark flips itself; the comparewatch- plane row flips with it. TressFX mirrored as the reference implementation; our current state is ribbon cards from generated guides -- the strand-system rung is the fork-chain leg 1. | ○ | ◉ | ◉ | ● | ● |
Marschner-class shadingOpen — watchingruntime/nx_game_page_emit.nx : hr_marschner_lobe, re-measured on every compare beat. Ship that symbol and this mark flips itself; the comparewatch- plane row flips with it. Marschner 2003 via Karis 2016 production reduction, mirrored; current hair shading is flat lock color. | ○ | ◉ | ◉ | ● | ● |
| FORK | |||||
Infinigen subsystem map live censusOpen — watchingruntime/nx_graphics_board.nx : if_fork_map, re-measured on every compare beat. Ship that symbol and this mark flips itself; the comparewatch- plane row flips with it. Only Infinigen itself holds column 1; the row measures OUR mapping coverage of their subsystems with sovereign twins, references and gates. | ○ | ● | ○ | ○ | ○ |
Debt register
| Id | Sev | What it is | Unblock |
|---|---|---|---|
| charjudge-blindnot in this scope | 5 | The single referee is hue- and semantics-blind in both directions: green clay scored 257 while a photoreal-class splat render scored 0, its smoothness read as no-detail. Every fidelity claim is ungradeable through it alone. | CLOSED 2026-08-19: cjc_panel landed and ADMITTED at FP=0 FN=0 on labelled corpus v1 -- fidelity claims now ride the panel's verdicts; residual inside GR1: the VLM lens is unobservable until a sovereign VLM oracle exists, and the floors need the second-corpus retest. |
| avatar-v1-qualitynot in this scope | 5 | The operator verdict on the v1 splat portrait stands recorded; causes measured and mechanical: un-searched body knobs (the 5-knob search is FLAT -- the form lives in the 16 canon ring rows), no face composite, no hair, and a splat radius that melts features. | GR2: canon-mode search plus composing the gated face pipeline plus hair volume; promote and register the organ out of the build lane. |
| 1786646407not in this scope | 5 | Arena overrun in the shared image-decode path on webp and one jpg: a write past the allocation while pixels still render -- the harmless-while-written, fatal-when-read class. | Fix the decode lib, then re-run the compositor smoke until the canary lines stop appearing. |
| magic-anatomynot in this scope | 4 | Rule-11 honest counts at threshold 2: the body factory holds 510 sites over 107 distinct values and the anatomy renderer 461 over 88. Most are structural; the anthropometric band edges 600 and 764 per-mille are POLICY hiding as arithmetic. | Hoist the band edges into the canon table as data, in the body lane with its own gates; prove neutrality by byte-identical rebuild. |
| advances-403-checknot in this scope | 2 | www.realtimerendering.com serves 403 to our client while the kesen and advances subdomains serve; a mirror that does not verify per-root reachability would publish false coverage. | GR7 prints reachable, mirrored and dead per root with the partition reconciled, exactly like the banked Ke-Sen run (236 reachable, 235 mirrored, 1 dead). |
Risk register
| Risk | Likelihood x impact | Mitigation |
|---|---|---|
| The referee gap gates everything: shipping fidelity rungs before GR1 reproduces the gumby-through-a-green-board failure | certain x high | GR1 first; flip bars pre-declared before experiments; the operator eyeball stays a named lens on the panel. |
| The box is saturated -- swap high, edge 503s under load -- and render sweeps and regens contend with it | likely x medium | Ask the spend gate before sweeps, batch off-peak, publish box receipts beside results, and stop writing when the edge 503s. |
| A watch contract flips on symbol existence, so a stub could read green | possible x high | The done-rule proven by a gate is the bar; the flip is the receipt -- standing law on every rung above. |
| Infinigen is a moving bar: 2.0 alpha is shipping the indoor frontier right now | likely x medium | Bars carry their sources and dates; the frontier watch topics aimed at this arc on 2026-08-15 feed the library on the beat. |
| Estimates rest on one graphics calibration point (R21) | certain x medium | Recalibrate on every landed rung and publish actual against estimate; the mangagen corroboration is banked. |
On these two registers. Rows are declared in the domain's plan file and carry the debt id, which is the join key back to the sovereign debt plane — that plane, not this page, is the authority on state. Reconciling them automatically (the regen reading the plane and refreshing these rows) is a named, owed rung; until it lands, treat an id here as a pointer to look up, not a status to trust.
Rendered evidence
These are the artifacts, not illustrations. Every image below was probed on disk at emit time and carries its measured byte size plus the organ that regenerates it; an absent file is named as absent and never linked. Re-run the producer and the page inherits the fresh render on the next compare beat.
Honest verdict. The coverage above is capability presence measured against source — not depth, scale, or polish, where mature rivals may lead. Exceeds are claimed only where a mechanism backs them. Every open gap is a watch contract: it names the organ and symbol that closes it, and this page flips the cell itself when that workstream ships.
Person · product · place — not yet measured for this domain
Every compare carries this layer. Declare
knowledge/compare/graphics.ppp (rows surface|nishi or c1..c4|label|url|connect naming OUR live surface and each rival's front door), run nx_ppp_probe domain graphics, and this section fills itself on the next beat: the same ruler on both sides — privacy and CX (third-party hosts, tracker classes, cookies, security headers), design and longevity (design hygiene, computed WCAG contrast, render-blocking resources, unsized media, script weight, theme and motion queries), findability (landmarks, skip link, on-site search, breadcrumb, headings, internal links).References
Beyond a link list. Every reference below resolves twice — the publisher's copy and, where banked, the estate's own non-rottable library mirror with a content pin — and carries its evidence class plus the exact claim on this page it grounds. Keyed marks like [key] in the matrix notes jump here. A dash means honestly absent, never assumed.
- [mei2007] Mei, Decaudin, Hu. Fast Hydraulic Erosion Simulation and Visualization on GPU. Pacific Graphics 2007. publisher · accessed 2026-08-18 · published-paperGrounds: EH1's hydraulic lane: the pipe-model shallow-water erosion this rung implements integer-exact, with mass conservation promoted from float hope to gate tooth.
- [musgrave1989] Musgrave, Kolb, Mace. The Synthesis and Rendering of Eroded Fractal Terrains. SIGGRAPH 1989. publisher · accessed 2026-08-18 · published-paperGrounds: EH1's thermal-talus lane and the founding statement that noise terrain without erosion reads wrong -- the classical bar the game tier meets.
- [terraindiff26] InfiniteDiffusion / Terrain Diffusion: training-free spatially-infinite diffusion with seed-consistency and constant-time random access; a learned real-time infinite terrain generator with a procedural-noise interface. SIGGRAPH 2026 (top-10 attention). publisher · read in our library
knowledge/library/gfx_sig2026.html· pinh85877ffc6f9b28d2e3a303e48585e7dd20dc397b1fc0c9a5d93a0ecc40645c2f· accessed 2026-08-18 · published-paperGrounds: THE 2026 CEILING for the beauty tier: the frontier moved from hand-tuned noise+sim to LEARNED generators that keep procedural properties. Our integer angle rides the nofloat ML stack; any beauty-tier terrain claim grades against this class, not against classical noise. - [metahuman-dna] Epic Games. MetaHuman DNA, Rig Logic and pipeline reference: head 24K-60K tris at LOD0, ~700-800 joints, 710 FACS-driven expressions, 8K source textures streamed 4K, per-vertex pore-level deformation LOD0-only, strand grooms at 3-6ms/character on 3070-class GPUs. publisher · accessed 2026-08-18 · vendor-docGrounds: THE CHARACTER DETAIL FLOOR the charsim ladder must budget against -- these numbers are the done-rule bar, and every earlier smaller budget on this board was an understatement corrected 2026-08-18.
- [metahuman-ref0] Epic Games. MetaHuman landing-page hero still (social-share render of a MetaHuman head, 1200x675 PNG), fetched through the sovereign fetcher 2026-08-19 and banked as the far-bar GOOD reference for the referee panel; truecolor twin cmp_graphics_metahuman_ref0_rgb.png via nx_charjudge_census topng. publisher · read in our library
knowledge/fetched/cmp_graphics_metahuman_ref0.png· pinh598d5eaedd10f0a65d9d914c884bc9d641175c224d6b913235cc4c5cd4520c55· accessed 2026-08-19 · vendor-docGrounds: REFEREE: multi-lens judge panel that can see beauty - [metahuman-ref1] Epic Games. MetaHuman landing-page hero still (cinematically graded MetaHuman face, 960x540 JPEG from the CMS asset pipeline), fetched 2026-08-19 and banked as the second far-bar GOOD reference; truecolor twin cmp_graphics_metahuman_ref1_rgb.png. The second-corpus retest on these two stills refuted the Kovac daylight skin floor (graded skin reads 0-6 permil) while the percept lens separated at fp=0 fn=0. publisher · read in our library
knowledge/fetched/cmp_graphics_metahuman_ref1.img· pinh36f796370a2e8795b0085296903754b1eaff82698a0b0ed9251b349a8a95d32e· accessed 2026-08-19 · vendor-docGrounds: REFEREE: multi-lens judge panel that can see beauty - [webgpu-status26] W3C GPU for the Web Community Group. WebGPU Implementation Status (living page); Safari 26 / Firefox 147 ship notes. Accessed August 2026. publisher · accessed 2026-08-18 · vendor-docGrounds: The DELIVERY tier ladder and its coverage claim: ~82 percent global WebGPU support as of 2026-08 (Chrome/Edge 113+, Safari 26 default-on, Firefox 147 Windows/ARM64-macOS with Linux through 2026) -- why WebGL2 remains the second tier and sovereign wasm the floor.
- [webgpu-roadmap26] Chrome GPU team. What's next for WebGPU (subgroups shipping behind flags, bindless in proposal via texture_and_sampler_let, no ray-tracing extension in spec). Accessed August 2026. publisher · accessed 2026-08-18 · vendor-docGrounds: The WebGPU-tier contract scope: RT stays compute-shader on the web in 2026; subgroups and bindless are the watched upgrades, not assumptions.
- [wasm3-state26] WebAssembly 3.0 in browsers: WasmGC, Memory64, relaxed SIMD, threads (COOP/COEP), tail calls -- stable in Chrome 119+/Firefox 120+/Safari 18.2+. Accessed August 2026. publisher · accessed 2026-08-18 · vendor-docGrounds: The sovereign-wasm floor's headroom: our MVP-profile emission can adopt SIMD/threads/Memory64 wherever the vm-gate proves equivalence -- the 85-95 percent-of-native band the delivery rows lean on.
- [infinigen23] Raistrick, Lipson, Ma et al. Infinite Photorealistic Worlds Using Procedural Generation (Infinigen). CVPR 2023. publisher · accessed 2026-08-17 · source-readGrounds: The Infinigen column codes and the single-surface-vs-layers differentiator: read from princeton-vl source (BSD-3), banked on the graphics board -- NURBS part lofting, genome part interpolation, per-species material node groups.
- [infinigen-indoors24] Raistrick, Mei, Kayan et al. Infinigen Indoors: Photorealistic Indoor Scenes Using Procedural Generation. CVPR 2024. publisher · accessed 2026-08-17 · published-paperGrounds: The ig_room_solve rung: constraint-based room solving with furniture placement is their published method; our contract is the same capability through our engine, refusing unsatisfiable sets.
- [infinigen-docs] Infinigen 2.0 documentation (Princeton Vision and Learning Lab). publisher · accessed 2026-08-17 · vendor-docGrounds: The 2.0 alpha capability read: controllable APIs, indoor materials and simplified rooms, procfunc composable generators, ground-truth annotation pipeline, USD/OBJ/FBX and MuJoCo export.
- [infinigen-flying26] Infinigen team. 2026-08 flying-indoor-preview dataset (17,500 video rows, 331 GB, BSD-3, under construction). publisher · accessed 2026-08-17 · datasetGrounds: The ws_groundtruth_emit contract: their annotation pipeline is why this dataset exists; ours makes the estate a dataset factory on the same axis.
- [kerbl3dgs] Kerbl, Kopanas, Leimkuehler, Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM TOG 42(4), SIGGRAPH 2023. publisher · read in our library
knowledge/library/gfx_sig2023.html· pinh3cc33d09097d5970c376313c7c0a89efbb2b09dfa4411bf053b4c36d22d2d0ba· accessed 2026-08-17 · published-paperGrounds: The SPLAT rows: gs_render_aniso implements the representation sovereignly in integer; gs_color_grad_step is the first tooth of the trained-fit loop the paper defines. - [splattingavatar24] Shao, Wang, Zhu et al. SplattingAvatar: Realistic Real-Time Human Avatars with Mesh-Embedded Gaussian Splatting. CVPR 2024. publisher · published-paperGrounds: The gb_bind hybrid: splats bound barycentrically to mesh triangles, mesh deforms and splats follow -- the operator-ordered outside-splat inside-mesh architecture.
- [penner2011] Penner, Borshukov. Pre-Integrated Skin Shading. GPU Pro 2, 2011 (SIGGRAPH course lineage). publisher · accessed 2026-08-17 · published-courseGrounds: The sr_sss_preint contract and the shipped wrap-SSS floors: diffusion widens the terminator per channel, red furthest -- the T66 equality tooth pins exactly the consequences no tuning can fake.
- [walter2007] Walter, Marschner, Li, Torrance. Microfacet Models for Refraction through Rough Surfaces. EGSR 2007. publisher · published-paperGrounds: The gp_ggx row: GGX/Trowbridge-Reitz D with Smith G and Schlick F is this paper's model, implemented integer-exact in the sovereign renderer.
- [karis2021] Karis, Stubbe, Wihlidal. Nanite: A Deep Dive. SIGGRAPH 2021, Advances in Real-Time Rendering in Games. publisher · read in our library
knowledge/library/gfx_adv_s2021_index.html· pinhb7941aaaa5a8921ff2c2fb13d8b8d6dada9e64642b696badab535d83524cd9e1· accessed 2026-08-17 · published-courseGrounds: The geometry-scale bar the retired hand page named; the s2021 Advances page carrying this talk is now mirrored in the library (the GR7 crawl, landed 2026-08-17). - [bitterli2020] Bitterli, Wyman, Pharr, Shirley, Lefohn, Jarosz. Spatiotemporal Reservoir Resampling for Real-Time Ray Tracing with Dynamic Direct Lighting (ReSTIR). ACM TOG 39(4), SIGGRAPH 2020. publisher · read in our library
knowledge/library/gfx_sig2020.html· pinh2f5975ded55073efa6ee13c3f33cbce9ca4f2965bf0644a3809a284ba05358bb· accessed 2026-08-17 · published-paperGrounds: The gp_bounce_gi contract names ReSTIR-class sampling as the bar beyond our cached-form-factor radiosity. - [chardon91] Chardon, Cretois, Hourseau. Skin colour typology and suntanning pathways. International Journal of Cosmetic Science 13, 1991. publisher · published-standardGrounds: The si_albedo_of_gene row: ITA (Individual Typology Angle) classes dark to very-light are this paper's typology; nx_skin_ita computes them in CIE Lab integer arithmetic.
- [bodyparts3d] Mitsuhashi, Fujieda, Tamura et al. BodyParts3D: 3D structure database for anatomical concepts. Nucleic Acids Research 37, 2009. CC BY-SA 2.1 JP. publisher · datasetGrounds: The oracle behind the BODY and ANATOMY rows: 1.26k meshes used as MEASURING STICK only, never shipped as product -- the product is the procedural generator measured against it.
- [kesen-mirror] Ke-Sen Huang's computer graphics paper index (SIGGRAPH, SIGGRAPH Asia, EG, I3D, EGSR, SCA, SGP, HPG, PG, 2000-2026), mirrored into the estate library. publisher · read in our library
knowledge/library/gfx_papers_index.txt· pinh1650826094a4b15ed0ac0877414162d2309cdcac4d41115ad292053ff0314aa5· accessed 2026-08-15 · mirrored-indexGrounds: The RESEARCH rows and every gfx_papers_index mirror path above: the fixed-point crawl proved link-completeness (reachable=236 mirrored=235 papers=9762, partition reconciled); one link was already dead at capture, which is why bytes beat bookmarks. - [advances-archive] Advances in Real-Time Rendering in Games, SIGGRAPH course series 2006-2026 (official archive). publisher · read in our library
knowledge/library/gfx_adv_s2026_index.html· accessed 2026-08-17 · published-courseGrounds: LANDED 2026-08-17: the archive is mirrored beside the Ke-Sen index under the gfx_adv_ prefix -- reachable=26 mirrored=25 dead=1 failed=0 truncated=0, partition RECONCILES (the one dead link had already rotted at capture, which is the argument for bytes over bookmarks). The mirror link here opens the newest year; the remaining tail is the extractor citing adv_ pages into the paper index. - [metahuman] Epic Games. MetaHuman Creator documentation. publisher · accessed 2026-08-17 · vendor-docGrounds: The Unreal Engine 5 column codes on character rows and the photoreal far bar; the north star (the inside-and-out twin) is graded against MetaHuman-class stills by the cjc_panel contract.
- [infinihuman25] Xue, Xie, Kostyrko, Pons-Moll. InfiniHuman: Infinite 3D Human Creation with Precise Control. arXiv:2510.11650, October 2025. InfiniHumanData: 111K identities annotated with text, multi-view images, clothing and body-shape parameters; InfiniHumanGen: diffusion-based controllable avatar generation. publisher · accessed 2026-08-18 · published-paperGrounds: The operator's realistic-human oracle ask (2026-08-18, an Infinigen-2.0 search): Infinigen ships NO human generator (source-read: its genome classes are creatures), so the human-realism bar on this axis is InfiniHuman-class controllable synthesis beside the MetaHuman stills -- the gs_avatar_genome leap and the cjc_panel referee grade against these external oracles, never our own eyeball.
- [kkguides] Community Koikatsu character-creation guide pack (KK Guides): character_creation_guide.pdf 4.5MB, Wubsy material-editor guides 1-4, shadow-texture guide, 60+ studio and shading items. pCloud public share, folder read 2026-08-18 via the publink API. publisher · accessed 2026-08-18 · published-courseGrounds: Operator-supplied oracle for the KK retarget lane (/compare/koikatsu, nx_kkfacts, nx_retarget_showcase) and the MATERIAL and skin rows: the pack names the material-editor and shadow-texture practices a KK-class card expects; our sovereign emit is measured against what the guides teach -- reference only, no third-party asset ever ships as product.
- [relooper-zakai11] Zakai. Emscripten: An LLVM-to-JavaScript Compiler. OOPSLA 2011 (SPLASH companion). The paper that names, specifies and proves the termination of the Relooper, whose three block kinds are Simple, Loop and Multiple. publisher · read in our library
knowledge/fetched/cmp_graphics_relooper-zakai11.pdf· pinhcfe4d5d4207254a34fc9f4937a622d2e7cbdc57aac4dbaa0e6ad0972006253ca· accessed 2026-08-23 · published-paperGrounds: DELIVERY: shaders compiled from NishiLang (nx_wgsl backend) -- the rung 2b structured control-flow recovery. Extracted with nx_pdf_text and quoted only from bytes read: a Loop block's Inner block is the one where "when execution reaches the end of that block, flow will return to the beginning" and which "will contain a conditional break defining where it is exited", which is exactly the S_LOOP plus S_BREAK form shl_region emits. The paper's label variable is introduced there because it "greatly simplifies the problem that the Relooper needs to solve" for multi-entry regions; our measured corpus has none, so we carry none and refuse by name instead. - [llvm-cfgstackify] LLVM Project. WebAssemblyCFGStackify.cpp, the WebAssembly backend's CFG stackification pass, which inserts BLOCK, LOOP, TRY and TRY_TABLE markers because scope boundaries serve as the labels for WebAssembly's control transfers. publisher · read in our library
knowledge/fetched/cmp_graphics_llvm-cfgstackify.cpp· pinh9bae82c88eb9ec5aba93cf7e9323da5a54b17ccb62cf3f8e7a3051e60345d413· accessed 2026-08-23 · source-readGrounds: DELIVERY: shaders compiled from NishiLang (nx_wgsl backend) -- the precondition that licenses a shape reducer instead of a full relooper. Read from the mirrored source: marker insertion "is sufficient to convert arbitrary CFGs into a form that works on WebAssembly, provided that all loops are single-entry". Every loop header in the measured shader corpus has exactly two predecessors, one preheader and one latch, so that precondition holds here; shl_region refuses BY NAME the moment a header appears with n_preds != 2. - [syn4d] Jiang, Lan, Luo, Deng, Lai, Sucar, Rupprecht, Laina, Larlus, Zheng, Vedaldi. Syn4D: A Multiview Synthetic 4D Dataset. ECCV 2026 Highlight. Visual Geometry Group University of Oxford, Nanyang Technological University, and Naver Labs Europe. publisher · read in our library
knowledge/fetched/cmp_graphics_syn4d.html· pinh11884d72fd639550e8f6828c96a8161aee786f82bda16f1a9b565afac6f7db6f· accessed 2026-08-25 · datasetGrounds: WORLD: ground-truth dataset emission -- RGB, depth, normals, semantics. WITNESSED in the pinned 42665-byte mirror and never through a summarising fetcher: their annotation pipeline renders dynamic scenes with a procedural multi-camera setup that jointly outputs RGB, depth, instance segmentation, optical flow, dense 3D tracks, and parametric human pose, and every scene carries per-pixel depth, multi-view camera trajectories, dense long-range 3D point tracks, and parametric SMPL-X human pose and shape. THIS IS THE 2026 PUBLISHED BAR THE ws_groundtruth_emit CONTRACT IS MEASURED AGAINST, and it is the same shape as our own emit verb because our renderer already holds depth and normals per pixel. Their stated differentiator is the ability to unproject any pixel into 3D at any frame and re-project it into any other camera, a cross-view and cross-time consistency property our single-camera snap does not yet claim -- named here as the gap rather than assumed away. Sits beside infinigen-flying26 as the second dataset-factory oracle on this board, and unlike Infinigen it is HUMAN-centred, which is what ties it to the charsim body rows. - [physrig] Zhang, Xu, Feng, Jampani, Ahuja. PhysRig -- Differentiable Physics-Based Skinning and Rigging Framework for Realistic Articulated Object Modeling. arXiv:2506.20936v2 cs.CV, 27 Jun 2025. University of Illinois Urbana Champaign and Stability AI. publisher · read in our library
knowledge/fetched/cmp_charsim_physrig.pdf· pinh744fc02485dfc52444abc672e445fff96ba284a36684f79ac830a14abbcd3f5d· accessed 2026-08-25 · published-paperGrounds: MOTION: DAMPING identified from video -- the parameter the field does not fit. Quoted ONLY from the pinned bytes via pdftotext and never through a summarising fetcher, because WebFetch on a PDF returns plausible sentences that are not in the document, measured twice on this estate. It embeds a rigid skeleton in a tetrahedral volume simulated as a soft body under continuum mechanics and introduces MATERIAL PROTOTYPES to shrink the learning space -- the same shape as our XPBD tet cage and the st_profile genome draw, which is why it is the right rival to name here. THE REASON IT BACKS THIS ROW IS WHAT IT DOES NOT DO: it optimises Young's modulus and Poisson ratio and NEVER damping, and damping is the parameter that governs jiggle decay and the one nx_dynaoracle identifies from video. - [softsmpl] Santesteban, Garces, Otaduy, Casas. SoftSMPL -- Data-driven Modeling of Nonlinear Soft-tissue Dynamics for Parametric Humans. Computer Graphics Forum, Proc. Eurographics, 2020. publisher · read in our library
knowledge/fetched/cmp_charsim_softsmpl.html· pinhce83469210b59198e67c539cffc9d2492d588110cbe0e1d5c18ae440e8ab5cef· accessed 2026-08-25 · published-paperGrounds: MOTION: one damped-oscillator identifier shared by the solver AND by video. The closest published prior art to this lane: soft-tissue dynamics as a function of body shape and motion for PARAMETRIC humans, which is the measure-then-generate loop this board is building. Its motion descriptor disentangles the pose representation by removing subject-specific features -- the same separation our gait DRIVE and tissue PLANT are deliberately kept apart by, and the reason a plant tuned into its own drive band reads as wrong. IT LEARNS FROM 4D SCANS, NOT FROM MONOCULAR VIDEO, so it does not close the video-identification gap this row claims and is cited as the bar rather than as the answer.
generated by nx_swcompare_matrix (sovereign NishiLang organ) from knowledge/compare/graphics.matrix · every Nishi cell verified against organ source at emit time · watch cells re-measured on every compare beat · zero JS, zero trackers