Graphics / Engine Convergence — the QA-loop comp

Status: FLAGGED (2026-07-12). This page is the standing comparison the workstream must work until the end user signs off SOTA. Judging is organ-run (nx_natstat / census gates / nx_ui_judge lineage), never hand-prose. Board: /world/foundation.

The comp — us vs the named field

capability	the SOTA bar (named)	us, organ/gate-measured	verdict
Render realism	MetaHuman / Infinigen photoreal (real-photo corpus LEVEL 776)	mean 421 across 7 renders (380→406→414→421 as Q-rungs land)	CLAY — gap 355, falling
Geometry scale	Nanite: 128-tri clusters, auto-LOD, streaming (Karis SIGGRAPH 2021)	generative LOD (zero stored clusters) + cluster decimation; capacity inversion proven	PARTIAL
Dynamic GI	Lumen: per-frame scene-reactive GI + reflections	cached form factors + incremental invalidation, 20.3x less work, bit-exact == rebuild	PARTIAL
Worldgen coherence	Minecraft-class staged pipeline: coherent hydrology, aquifers, structures	6-param staged pipeline + Q3 HYDROLOGY (downhill traces to the sea, monotone beds) + altitude lapse (snow 54% peaks vs 20% low)	PARTIAL — coherent, thin (no tributaries/structures)
Rigged humans	MetaHuman rig: photoreal face, groom, capture pipelines; SMPL scan-fit bodies	9-joint FK walker + curl hands + muscle-driven expressions; NO anthropometric/atlas correctness	PARTIAL — mechanisms only
Internal anatomy	ZygoteBody-class medical atlas (peel, correct shapes)	peel + 8 systems/119 parts + Q4: proportions MEASURED-CANON from the SDF (7.7 heads, 5/5 ratios in band vs banked anthropometry)	PARTIAL — proportions correct; organ shapes vs atlas still open
Globe-scale world	CesiumJS: WGS84 high-precision globe, 3D Tiles streaming, global terrain, 4D time-dynamic (banked 07-12)	R6 GENERATIVE STREAMING: unbounded world, bounded buffer, bit-identical revisit from seed (their tile servers vs our 8 bytes); no WGS84 frame yet	PARTIAL — streaming real, globe frame open
Creature generation	Infinigen: 5 tree-genome classes (Carnivore/Herbivore/Bird/Beetle/Fish), NURBS parts, fur, surface-only (banked 07-12)	ONE tree-walking generator, 3 plans (canon human / rabbit / 4-arm monster), live genes, surface-only default + medical-view internals THEY LACK	PARTIAL — fewer plans, no curved parts/fur yet
AI authoring (Orca-class)	Orca: describe a Minecraft mod → AI writes/compiles/installs it; AI ops read crash logs + auto-fix; server live in 30s (banked 07-13)	describe a creature OR a WORLD → built + intent-validated + deployed (nx_authorgen creatures 7/7 + worlds 6/6); anti-hallucination (refuses/reports unmet); Orca pieces exist as islands — integration is the gap	PARTIAL — 440/1000; real gap = wiring

Exceed claims — only with a WHY

Bit-exact incremental GI exceeds because integer determinism makes cache-vs-rebuild EQUALITY provable — float engines can only approximate-compare (mechanism + measured equality, nx_cornell_dyn_gate). Generative LOD exceeds Nanite for generated content because the generator IS the compression — zero stored cluster hierarchy (measured: 3 LOD tiers from one SDF, deterministic; capacity inversion). Anatomy presence is NOT claimed as exceed anymore — the reality audit demoted it to presence-without-correctness. "Sovereign" alone is never a why.

Organ-measured render reality (the clay numbers)

ORGAN-MEASURED RENDER REALITY -- graded by nx_natstat ns_assess (NSS/MSCN no-reference IQA, corpus-calibrated), NOT hand-prose. REFERENCE: real-photo corpus LEVEL mean 776 (min 600) / Z-Image AI target 676 (nx_natstat_corpus 2026-07-06). Bands: >=744 PHOTOREAL-GRADE, ~400-600 SEMI-REAL, below = CLAY territory.

  engine scene (R1)     LEVEL=482  [scale=540 mscn=482 kurt=1000 colour=1000 ai-fp=1000] 1040x380
  LOD scene (R2)        LEVEL=284  [scale=480 mscn=284 kurt=1000 colour=1000 ai-fp=1000] 1040x380
  dynamic GI (R3)       LEVEL=552  [scale=0 mscn=552 kurt=1000 colour=1000 ai-fp=1000] 1080x380
  hands (R4)            LEVEL=285  [scale=0 mscn=285 kurt=1000 colour=1000 ai-fp=1000] 900x320
  face exprs (R4)       LEVEL=456  [scale=0 mscn=456 kurt=1000 colour=1000 ai-fp=1000] 760x250
  staged world (R5)     LEVEL=433  [scale=280 mscn=433 kurt=1000 colour=1000 ai-fp=1000] 920x360
  being peel (DH)       LEVEL=354  [scale=0 mscn=354 kurt=1000 colour=1000 ai-fp=1000] 1248x368

MEAN 406 (min 284, max 552) vs real-photo corpus 776 / AI target 676. The gap to photoreal is 370 -- ORGAN-measured, tracked per quality rung Q1-Q7. Weakest axes across renders = scale-invariance + MSCN local-contrast (multi-scale self-similar detail; matches the engine gap-map of 2026-07-06). These are exactly what Q1 texture + Q2 shadows attack.

The roadmap this workstream is working (quality-first)

rung	what	the organ that must move	state
Q1	Texture the world (multi-octave albedo, hue octave)	ns_assess MSCN + scale axes rise	DONE 07-12 (engine scene 386→482)
Q2	Sun shadows (world-space ortho shadow map)	causal darkening + level	DONE 07-12 (1449 darkened / 0 brightened)
Q3	Hydrology: downhill rivers + altitude lapse	worldgen-coherence axis	DONE 07-12 (5/7 rivers reach the sea; engine 314→342)
Q4	Anatomy reference pass (anthropometry canon)	anatomical-correctness axis	DONE 07-12 (5/5 ratios; reality 150→200; atlas shapes = residual)
Q5	Foot planting / ground contact	motion-realism axis	DONE 07-12 (28-unit float closed to ≤4, causal)
Q6	Face sculpt (eyes, lip seam, skull proportions)	face-realism axis	DONE 07-12 (thirds 7% dev; eyes gloss-proud; reality 200→250)
Q7	Judge growth: every operator-seen miss becomes a new organ axis	the judge itself	standing

Check-ins & buy-off

date	check-in
2026-07-13	#11 — authoring content type 2: WORLDS. "a snowy mountainous world" / "an ocean world with islands" / "a flat hot desert" → the pipeline generates the measured character from adjectives (snow 161→591, ocean 295→818; causal, deterministic; non-worlds refused). nx_authorworld_gate 6/6.
2026-07-13	#10 — ORCA-class AI authoring (operator ref orcaclient.com): describe-a-creature → build → validate → deploy (nx_authorgen 7/7, anti-hallucination proven); AI-authoring census 440/1000. Scout finding: Orca pieces (forge codegen+repair, self-heal, golive hosting, app-publish registry, Qwen :11434 seat) exist as ISLANDS — roadmap is INTEGRATION not build.
2026-07-12	#9 — PHYSICS rung 1 (fall/bounce/rest on the world, contact-exact, bit-deterministic replay; engine 346→373) + LOD scene retextured (judge 284→331; mean 421).
2026-07-12	#8 — Q6 face sculpt: eyes/seam/canon-thirds measured (7% dev, eye-line 5%); reality 200→250; + nx_cc scope-slot compiler bug characterized.
2026-07-12	#7 — operator: "awful clay monsters, no anatomy intelligence" → smooth-union skin (joints flow; all gates re-verified) + CREATURE GENOME (one generator: human/rabbit/monster from data; live genes; surface-only research default, internals opt-in).
2026-07-12	#6 — Q5 foot planting SHIPPED: 28-unit mid-stride float measured then closed (≤4, all poses, causal band proof); residual weight-shift/heel-toe/IK.
2026-07-12	#5 — R6 generative streaming SHIPPED: 10-station journey, world > caps, station ≤18k, bit-identical revisit, real eviction; engine 320→346.
2026-07-12	#4 — Q4 anthropometry SHIPPED: being taped from its SDF, body FIXED to canon (5/5 ratios); reality anatomy GAP→PARTIAL 150→200. CesiumJS banked → globe-scale axis added as GAP (engine 342→320, instrument widened honestly).
2026-07-12	#3 — Q3 hydrology SHIPPED: steepest-descent rivers (monotone beds, 5/7 reach the sea), altitude-lapse snow (54% vs 20%, measured), alpine relief; engine coherence GAP→PARTIAL, 314→342.
2026-07-12	#2 — Q1 texture + Q2 shadows SHIPPED; organ re-grade mean 380→406 (engine scene 386→482, world 345→433); still clay vs 776.
2026-07-12	#1 — reality audit ran; organ grader wired (mean 380 vs photo 776); presence-inflation corrected: engine 416→314, digital-human 294→258; reality census 150 with inflation guard. Next check-in after Q1+Q2.

BUY-OFF: PENDING OPERATOR. When reviewed: misses the operator sees that the organs did not flag become NEW JUDGE AXES (native capability growth), then the workstream is REGRADED. This loop repeats until the operator signs off SOTA.