Character Sim and Breeding Games

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 Koikatsu and Honey Select 2 and Breeders of the Nephelym and Tropical Monster Girls -- the character-maker plus breeding-sim field. Measured 2026-08-14: nx_gamebench puts the GENERIC game substrate at 980 permil (25 HAVE of 26 capabilities), while nx_vsbench puts the CHARACTER-FIDELITY head-to-head at beat_index 0 permil over 10 cross-class cells. The distance between those two numbers is the whole roadmap. Only Breeders sits on an existing Nishi ruler; the other three titles are measured here for the first time.

Layer 1 · Executive

Where we are. Measured 2026-08-14 against live rulers, not recalled. The generic game substrate is strong and gated: nx_gamebench reports 26 capabilities at 980 permil, with save/load, rpg progression, entity store, world sim, plot generation, loot tables, pathfinding, audio and input abstraction all certified composable parts. The skin pipeline is further along than the written record suggested -- a bone-anchored UV unwrap plus a per-region albedo atlas with per-fragment sampling LANDED on 2026-08-12, nx_nxa_texbake_gate 11 of 11 GREEN with published artifacts, closing the 2026-08-03 finding that texturing was impossible rather than merely hard. So the deficit is not mostly capability, it is FIDELITY and REACH: the atlas is proven in the core renderer and absent from the shipped world shader, the palette is region-flat rather than genome-driven, and on the one measured head-to-head ruler all ten cells are cross-class MIXED at beat_index 0 permil. Three of the four titles named here sit on no Nishi ruler at all. The agent lane is the inverse case -- a memory stream, an adversarial reflection primitive and a structurally-winnable planner all exist as separate organs, and nothing has ever connected them into a loop.

Where we need to go. Meet the field on the two loops it is built from -- the creator loop (make a character, dress it, pose it, share it as a card) and the breeding loop (inherit, transform, raise) -- then exceed it on the axes this field structurally cannot hold: a character that is a SEED rather than an asset pack, an adult invariant that fails closed at the compiler rather than by policy, and emergent agents with memory, reflection and planning that not one of the four titles ships. Fidelity is a climb with a stated cost, never a claim; the agent lane is graded on outside benchmarks rather than on our own scoring.

The unit. 1 u = one measured session-leg. Calibration: the sibling mangagen lane took its panel compositor from existing substrate to shipped, registered and live-verified in a single leg on 2026-08-13. Estimates recalibrate as each rung lands.

Cost to stop reading as clay: 10.5 u. Through milestone M2. The first half-unit is the cheapest rung on the whole board, because the UV and albedo ladder already landed gate-green; what remains is driving it from the genome and getting it into the shipped world shader rather than only the lab renderer.

Cost to the first honest EXCEED: 11.5 u. R21. Today every head-to-head cell is cross-class MIXED by rule, so no perceptual win can be minted until our character render is the same class as theirs. That rule is deliberate -- it was added after a statistical win over a 3D render was correctly retracted by the operator.

Cost to the uncontested lane: 27 u. Milestone M5. Emergent agents are the one axis not a single one of these four titles contests. It is also the only lane where meeting the field is not the goal, because the field is not on it.

Cost to full parity plus exceed: 32.5 u. Everything below. Paired interaction animation and the VR headset runtime are the long poles at 3 u each and are genuine greenfield; every other rung composes something already certified.

Research bar. VaM asset corpus floor is measured on complete-skin texture floor, read from real image headers across 23 packages rather than from package names. Theirs: 4 UV regions times 4 PBR maps = 16 maps per skin, 4096px floor and 8192px premium, morphs 26 floor and 84 median and 310 high. Ours: our SKIN rows: region count already met exactly (104 bones partitioned face 13, torso 14, limbs 74, gens 3), albedo only, 2048px proven -- map count and resolution are the gap.

Research bar. nx_vsbench head-to-head board is measured on our frame scored against each title's own published render on one instrument at one scale. Theirs: Breeders 789, Carnal 845, MGI 905, VaM 907, DIK 913, Operation Lovecraft 990, and Mad Island, VR Hot and CM3D2 all saturated at 1000. Ours: ours 875 on the M2D frame and 1000-saturated on Craft -- reported MIXED and never EXCEED, because all ten pairings are cross-class.

Research bar. Frame-critic caricature control is measured on what a zero-art deterministic random mosaic scores on our own frame critic -- the calibration that convicts the scale. Theirs: 628 of 1000, graded RICH. Ours: every visual number on this page is a STRUCTURE statistic and not a quality rank; the commercial band 789-1000 overlaps the band reachable by coloured squares.

Research bar. OpenAI multi-agent hide and seek is measured on whether unscripted tool use emerges from simple physics primitives over long training. Theirs: seekers learned to ride ramps over walls; hiders countered by locking ramps away before the round began -- emergence with no scripted key. Ours: R15 and R17: nx_worldsim and nx_worldpipe already supply tools rather than keys, but no agent yet uses them [@emergent-tool-use19].

Research bar. Google Game Arena Werewolf is measured on negotiation, strategic deception and role-conditioned behaviour in hidden-role social deduction. Theirs: a published multi-model leaderboard protocol. Ours: R20 runs their protocol unchanged and publishes our row beside theirs [@game-arena-werewolf].

Research bar. Emergence World long-horizon autonomy is measured on behavioural drift, cross-model contamination and safety erosion over weeks rather than minutes. Theirs: explicit prohibitions erode under perceived structural failure; societal outcomes diverge sharply by underlying model. Ours: R19 -- and our compiler-enforced invariant is precisely the control case their prompt-level prohibitions lack [@emergence-world].

Research bar. SecretWaifu AI-companion reach is measured on what a shipping real-time AI-companion product claims on its own homepage, and the delivery surfaces it reaches -- the operator named this title on 2026-08-22 as a thing to be able to make. Theirs: a cinematic AI girlfriend experience with expressive characters, live conversation, browser play, Windows builds, PCVR, and Oculus Quest support [@swaifu]. Ours: R11 cm_npc_turn and R12 sr_vr_present are the rungs it contests directly, and R9 fa_expr_morph is what expressive characters means. ITS EXISTENCE REFUTES the nx_vsbench advantage hypothesis that not one reference title contests freeform in-game NPC dialogue -- that axis carries adult_titles_needing 0 today, which is exactly why R11 ranks 15 of 19 and sits in M5 at 27 u. A rung is cheap to reorder and a wrong deficit column is not, so the ruler is corrected before the order is read.

Research bar. Reference rig corpus width is measured on whether the reference corpus can express a RANGE at all, measured by nx_rigfloor over the full population rather than by file count. Theirs: the 11 rigged .nxa under sites/nishifamily/world are BYTE-IDENTICAL in geometry -- verts=14164 tris=28092 joints=104 on every one, one mesh with 11 animation variants, and the 12th carries no SKEL section at all; nx_rigfloor returns axes_with_width=0 and refuses to write a floor. Ours: A GENERATOR CANNOT BE BENCHMARKED AGAINST A CORPUS OF ONE -- the question is undefined rather than hard. This is a CORPUS defect and not an organ defect: the instrument is shipped, gate-green, and correctly abstains. R22 is the work that resolves it by turning donors into rig references.

Research bar. Donor-derived asset floor is measured on the game-ready floor as a MEASUREMENT of the donor population, replacing the five hand-picked bounds nx_asset_floor_gate carried (AF_VERT_MIN 4000, AF_TRI_MIN 8000, AF_TRI_MAX 120000, AF_JOINT_MIN 40, AF_MIDLINE 3000). Theirs: derived over 9 DISTINCT donors with corpus_complete=1 and axes_available=3 axes_with_width=3: tris 4033..120704, parts 1..9, tris_per_part 1477..120704, method=observed-range-full-population. Ours: nx_rigfloor LIVE and nx_rigfloor_gate GREEN 18 of 18 with three bite pairs; NOT a second ruler -- nx_asset_floor_gate stays the one ruler and reads the conf, so producer and consumer cannot disagree.

Research bar. Ingest round-trip fidelity is measured on what our own glTF ingest LOSES, measured as a SELF-ORACLE needing no external ground truth and no licensing decision -- the only oracle available while the reference corpus is degenerate. Theirs: attempted 16 = IDENTICAL 10 + LOSSY 6 + UNOBSERVABLE 0, summing exactly; vertex positions survive BIT-EXACT on all 16 (triangles_with_differing_positions=0) while named-part segmentation is a TOTAL loss on the 4 rigged donors, 9 layers collapsing to 1, byte arithmetic reconciling at 1426008-1425816=192=8x24 dropped layer records. Ours: R22. nx_gltf2mesh reads JOINTS_0 as a NAME KEY only (zero WEIGHTS_0, zero inverseBindMatrices, zero animations) and nx_mesh2glb emits no skins or JOINTS_0 at all, so NEITHER LEG CARRIES A RIG and the pipeline is form-only end to end.

Research bar. Rig floor (NXANIM01) is measured on the first RIG floor: a joints band derived by nx_rigfloor from the three distinct rigged assets the estate holds (ref9d plus the two glTF donors carried to NXANIM01 by R22), the full rigged population not a sample. Theirs: MetaHuman LOD0 713 joints incl facial (VENDOR-DOC, Epic HTML LOD table); VRM 1.0 15 required humanoid bones (SPEC). Ours: verts 6728..14164 tris 13444..28092 joints 9..104 tris_per_vert 1983..2000 permil verts_per_joint 136..908; axes_with_width=5 of 5; n_read=3 corpus_complete=1.

Research bar. Ship texture resolution (derived) is measured on the texture resolution we ship at, DERIVED as the largest MEASURED TEXM rung whose bytes times cast_size fits a CITED texture streaming pool -- never a picked number; tex_res_floor 4096 stays the measured bar it is compared against. Theirs: Epic AndroidEngine.ini DefaultMemoryBucket_MinGB=4 plus DeviceProfiles r.Streaming.PoolSize=180 MB for the Default bucket (VENDOR-DOC, mirrored and pinned h208f2e30...); Meta Quest 3 app memory kill limit 5.75 GiB (VENDOR-DOC, used as a bound only). Ours: floor=4096 ship=2048: pool 188743680 B over cast 12 -- 4096 needs 378767808 and does not fit, 2048 needs 69295680 and fits; named residual: the ladder is storage bytes, GPU-resident at 2048 is 64 MiB per character uncompressed and needs hardware texture compression.

Research bar. Rigged round-trip fidelity is measured on whether the RIG survives a full nishi to glTF to nishi loop through both legs (nx_mesh2glb container-dispatch export and R22 nx_gltf2mesh .nxa import), read back by an independent reader (nx_asset_floor_gate) and scored against the rig floor. Theirs: ideal: section wordlens identical and joint count unchanged. Ours: ref9d.nxa to glb (805456 B joints=104 skinned=1) to nxa: VERT:42493 TRIS:84277 SKEL:833 SKIN:113313 byte-identical to the shipped asset, joints=104, weight_sum_worst_dev_q12=0, bind_rot_worst_dev_q12=0, IN-BAND 5 of 5 axesbar.

26 of 48 capabilities measured|3 of them measured exceeds|22 open|coverage 541/1000|adoption 19 full / 7 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=1787883527 domain=charsim target_version=0.1 rungs=25 done=5 open=20 finish=0 ranker=nx_dr_ocm

#	Stage	Rung	Priority	Derivation
#1	0.1	Soft-body bound to skin (R7) dyna_bind_skin	6000	v=15 m=2 c=5
#2	0.1	Cloth and hair on the same solver (R8) dyna_cloth	2600	v=13 m=1 c=5
#3	later	Semantic recall over the memory stream (R14) em_recall_dense	2400	v=12 m=1 c=5
#4	later	Character creator surface (R1) cc_slider_ui	1500	v=15 m=2 c=20
#5	later	Portable character card (R2) cd_card_emit	900	v=9 m=1 c=10
#6	later	Feet adapt to ground and heels (R23) ap_foot_ground	866	v=13 m=1 c=15
#7	later	Paired interaction animation (R3) ap_paired_clip	833	v=25 m=1 c=30
#8	later	Observation engine (R15) obs_tick	700	v=7 m=1 c=10
#9	later	Dowson donor through the bake lane (R24) f2n_groom_bake	650	v=13 m=1 c=20
#10	later	Relationship and schedule sim (R10) rs_schedule_tick	600	v=9 m=1 c=15
#11	later	Reflection loop bound to an agent (R16) dr_reflect_agent	600	v=6 m=1 c=10
#12	later	Graph-state evaluator from play logs (R18) sv_arc_derive	300	v=3 m=1 c=10
#13	later	Intent oracle (R17) no_intent_exec	266	v=4 m=1 c=15
#14	later	VR headset runtime (R12) sr_vr_present	200	v=3 m=2 c=30
#15	later	Agent dialogue turn inside the loop (R11) cm_npc_turn	150	v=3 m=1 c=20
#16	later	Long-horizon drift harness (R19) eb_longrun	142	v=1 m=1 c=7
#17	later	Hidden-role social-deduction arena (R20) eb_werewolf	142	v=1 m=1 c=7
#18	later	Single illustration lifted to a 3D head (R13) p3d_lift	80	v=1 m=2 c=25
#19	later	First same-class head-to-head cell (R21) vs_ccls_same	0	v=0 m=3 c=10
#20	later	Rig-carrying glTF ingest (R22) g2_skin_weights	0	v=0 m=1 c=20

Critical path — contract, done-rule, executor, cost

Rung	Closes with	Definition of done (pre-declared)	Executor	Est.
Genome-driven albedo (R0)	ntb_albedo_genome	Skin tone and parameters read from the character genome instead of a region-flat palette, so a bred child inherits its own skin; distinct genomes must measure distinct on nx_charjudge, and a genome that changes nothing is a defect to rewrite	Organ	0.5 u
Full PBR map set (R4) after R0	ntb_bake_pbr	Albedo, specular, gloss and normal across all four regions with normal-map lighting in the renderer, and the 4096px floor actually proven rather than assumed from a knob; the judge must move, not merely the file count (est aligned to graphics.plan GR8, the consolidated owner, 2026-08-23)	Organ	1 u
Atlas sampling in the world NPC shader (R6) after R0	cw_npc_atlas	The craft-world NPC wasm shader samples the same atlas the core renderer already samples, so textured skin appears in the shipped world and not only in the lab	Organ	1 u
Character creator surface (R1) after R0	cc_slider_ui	Player-driven sliders re-emit the body live; every knob is an existing nx_body_proc parameter, and nx_charjudge grades the output rather than the eyeball	Organ	2 u
Portable character card (R2) after R1	cd_card_emit	A card is a STRING carrying the genome seed plus the adult invariant; it round-trips bit-exact through gs_save and a minor card REFUSES at the compiler with the buffer left untouched	Organ	1 u
Subsurface skin shader (R5) after R4	sr_sss_preint	Pre-integrated SSS bound to the baked normal map, with a red-team control proving the shader does more than tint the image warm (est aligned to graphics.plan GR9, the consolidated owner, 2026-08-23)	Organ	1 u
Soft-body bound to skin (R7) after R0	dyna_bind_skin	The gate-proven solver drives actual skin vertices; the stale 23192B promoted binary is refreshed past its 28747B build and the DYNA section round-trips through the asset (est aligned to graphics.plan GR13, the consolidated owner, 2026-08-23)	Organ	1 u
Cloth and hair on the same solver (R8) after R7	dyna_cloth	Wardrobe layers and hair strands driven by dyna_bind_skin, proven to be the same solver rather than a second system	Organ	1 u
First same-class head-to-head cell (R21) after R4,R5	vs_ccls_same	Our real-time 3D character render scored against a reference title's at ccls=0, so an EXCEED can be minted honestly instead of reported MIXED. The rung that makes the beat index mean anything	Organ	1 u
Paired interaction animation (R3)	ap_paired_clip	Two skeletons under contact constraints with no interpenetration, graded by the nx_clipcheck ruler that already exists and bite-proven in both directions. The defining capability of all four titles	Organ	3 u
Facial expression morphs (R9) after R4	fa_expr_morph	FACS-style expression axes on the emitted mesh; distinct expressions must measure distinct, and an axis that changes nothing is a defect to rewrite, never to delete (est aligned to graphics.plan GR12, the consolidated owner, 2026-08-23)	Organ	2 u
Relationship and schedule sim (R10)	rs_schedule_tick	NPCs run a day arc composing the certified worldsim and charmem parts; the loop replays deterministically and saves bit-exact	Organ	1.5 u
VR headset runtime (R12) after R5	sr_vr_present	A headset present-loop on our own renderer at a stated frame budget, no WebXR; stereo correctness is already proven, the present path is not	Organ	3 u
Semantic recall over the memory stream (R14)	em_recall_dense	em_recall's lexical ranker swapped for the trained dense model already live in the search lane; a pre-declared accept rule -- it must beat the lexical ranker on a held-out recall set or it is not wired	Local model	0.5 u
Agent dialogue turn inside the loop (R11) after R14	cm_npc_turn	Freeform per-NPC dialogue with latency measured INSIDE the game loop before anything is built on top of it. Multi-NPC cost is measured, not assumed	Local model	2 u
Observation engine (R15) after R14	obs_tick	World events become agent-legible percepts written to the memory stream; the partition of events to percepts is printed and must sum, so a silently dropped event cannot pass	Organ	1 u
Reflection loop bound to an agent (R16) after R15	dr_reflect_agent	nx_dr_refute's decorrelated N-verifier voting applied to an NPC's own memories so the agent revises its beliefs; a planted false belief must be killed and a true one must survive	Organ	1 u
Intent oracle (R17) after R16	no_intent_exec	Natural-language agent intent translated into coordinate-level pathfinding and object interaction over nx_pathfind and nx_inventory; an unexecutable intent REFUSES loudly rather than silently no-opping	Organ	1.5 u
Graph-state evaluator from play logs (R18) after R17	sv_arc_derive	A narrative graph derived FROM logged play, detecting when an unscripted action created a genuinely new arc while the game state stayed consistent	Organ	1 u
Long-horizon drift harness (R19) after R18	eb_longrun	A multi-week deterministic run measuring behavioural drift and safety erosion. The pre-declared claim under test: our compiler-enforced adult invariant holds where a prompt-level prohibition erodes -- and because the run is bit-exact replayable, any divergence is bisectable to the tick	Organ	1.5 u
Hidden-role social-deduction arena (R20) after R11	eb_werewolf	The published Game Arena protocol run unchanged, our row published beside theirs. A grade-1 external referee for negotiation and deception, never our own scoring	Local model	1.5 u
Single illustration lifted to a 3D head (R13) after R4	p3d_lift	A PANIC3D-class lift from one illustration to a 3D head, scored on chamfer and F1-at-5mm against the AnimeRecon paired benchmark -- their metric family already matches our meshdist ruler, so the comparison is direct	Local model	2.5 u
Rig-carrying glTF ingest (R22)	g2_skin_weights	nx_gltf2mesh reads WEIGHTS_0, inverseBindMatrices and the joint hierarchy instead of using JOINTS_0 as a part-naming key, and nx_mesh2glb emits skins and JOINTS_0 back out. The done-rule is mechanical and already instrumented: nx_roundtrip must report IDENTICAL on the four rigged donors it currently reports LOSSY, with named-part segmentation surviving the trip. Until this lands, NXMSH2 structurally cannot carry a rig and no rig floor is derivable from any donor	Organ	2 u

Milestones

Milestone	Rungs	Cumulative
M0 · Skin reads as skin, in the shipped world	R0,R4,R6	3.5 u
M1 · Creator loop usable	R1,R2	6.5 u
M2 · Flesh, not clay	R5,R7,R8	10.5 u
M3 · First honest same-class EXCEED	R21	11.5 u
M4 · The genre loop complete	R3,R9,R10,R12	20 u
M5 · An emergent agent society	R14,R11,R15,R16,R17,R18	27 u
M6 · Benchmarked against the outside	R19,R20	30 u
M7 · Field parity plus the uncontested lanes	R13	32.5 u
M8 · Ingest carries the rig	R22	34.5 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	Koikatsu	Honey Select 2	Breeders	Tropical Monster Girls
CREATOR
body derived from anatomy numbers, not an artist mesh Measured: `sg_fk` exists in `runtime/_hdl_build/nx_skelgen.nx`, verified at emit. Rivals morph a hand-sculpted base mesh; ours is generated from anthropometric canon [ansur2] plus forensic tissue depth, so every proportion is a parameter rather than a sculpt. Honest split: adjustability is ahead, surface fidelity is far behind. THE CLOSEST PUBLISHED RIVAL TO THIS EXACT PRINCIPLE IS NOW CITED AND PINNED: Anny [anny] reaches the same conclusion we do -- anthropometric and World Health Organization statistics instead of 3D body scans, for the same three reasons of limited populations, unrealistic child-to-adult interpolation and biometric privacy -- yet it still starts from a MakeHuman base mesh of over 13,000 vertices and 163 bones, layers 564 artist-defined blendshapes over it, and deforms with blend skinning, which is the sculpt dependency this row claims to remove. It also covers infants to elders under Apache 2.0, which we do not, so it is both the anthropometric oracle to grade the canon against and a named coverage gap. Multi-HMR 2 [multihmr2] is the reverse-direction referee: it recovers pose and shape from a single camera with child-aware body modelling, so a generated body whose recovered mesh disagrees with the parameters that produced it is a mechanically measurable failure -- the cheapest grade-1 external check available to this row, and one we do not yet run. 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). adoption REGISTERED-DARK	●	◉	◉	●	●
pose legality checked by anatomy Measured: `sg_legal` exists in `runtime/_hdl_build/nx_skelgen.nx`, verified at emit. sg_legal answers whether a joint angle is anatomically legal BEFORE anything renders. No title in this field validates a pose; they clamp and hope. 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). adoption REGISTERED-DARK	●	○	○	○	○
slider surface (the Chara Maker) Open — watching `runtime/nx_charcreator.nx` : `cc_slider_ui`, re-measured on every compare beat. Ship that symbol and this mark flips itself; the `comparewatch-` plane row flips with it. R1 KEYSTONE gap watching cc_slider_ui. The knob space already exists as nx_body_proc parameters; what is missing is the live re-emit surface a player actually drives [cc5-reallusion]. watching cc_slider_ui	○	◉	◉	●	●
portable character card Open — watching `runtime/nx_charcard.nx` : `cd_card_emit`, re-measured on every compare beat. Ship that symbol and this mark flips itself; the `comparewatch-` plane row flips with it. R2 gap watching cd_card_emit. Cards are the sharing economy of the Illusion titles. Ours must carry the genome seed plus the adult invariant, which makes a character a STRING rather than an asset pack. watching cd_card_emit	○	◉	◉	○	○
read a Koikatsu card structurally Measured: `kk_probe` exists in `runtime/nx_kkfacts_lib.nx`, verified at emit. kk_probe censuses a real KK card -- portrait pixels, payload bytes, and the custom, coordinate, parameter, status and kkex blocks -- structure only, creative content deliberately unread. The interop reader exists; nothing yet consumes what it reports. Adoption: `LIB-WIRED importers=4 nonval=3` — fully adopted (top of its ladder).	●	◉	●	○	○
single illustration lifted to a 3D head Open — watching `runtime/nx_p3d_lift.nx` : `p3d_lift`, re-measured on every compare beat. Ship that symbol and this mark flips itself; the `comparewatch-` plane row flips with it. R13 gap watching p3d_lift, and an axis NO title in this field contests -- all four require you to build inside their maker. Donor research banked (PANIC3D CVPR2023 [panic3d23], VRoid 11.2k, AnimeRecon paired benchmark) with their chamfer and F1-at-5mm metrics in the SAME family as our meshdist ruler, so their numbers and ours compare directly. Operator directive 2026-08-06 named Koikatsu as the target. watching p3d_lift	○	○	○	○	○
SKIN
UV unwrap and a TEXC section in the asset Measured: `ntx_apply` exists in `runtime/nx_nxa_texc_lib.nx`, verified at emit. ntx_apply computes a bone-anchored unwrap from SKEL plus SKIN and writes TEXC, additive and idempotent. LANDED 2026-08-12, regions-from-anatomy exact (face 13, torso 14, limbs 74, gens 3 = 104, sums checked). This CLOSED the 2026-08-03 finding that our mesh carried no UVs and sampling any texture was impossible, not merely hard. Adoption: `LIB-WIRED importers=7 nonval=4` — fully adopted (top of its ladder).	●	◉	◉	◉	●
per-region albedo atlas bake and sampling Measured: `ntb_bake` exists in `runtime/nx_nxa_texbake_lib.nx`, verified at emit. Deterministic even-grain albedo atlas baked from the TEXC grid as SSOT, with per-fragment sampling proven load-bearing through the renderer's existing image path. nx_nxa_texbake_gate 11 of 11 GREEN; atlas and textured-view artifacts published and eyeballed at /world/anatomy_gpu. Resolutions 1024 and 2048 proven. Adoption: `LIB-WIRED importers=5 nonval=2` — fully adopted (top of its ladder).	●	◉	◉	◉	●
albedo driven by the character genome Measured: `ntb_albedo_genome` exists in `runtime/nx_nxa_texbake_lib.nx`, verified at emit. R0 gap watching ntb_albedo_genome (debt 1786549493 residue a). Today the palette is region-flat; skin tone and parameters must come from the genome so a bred child inherits its own skin. Rule 25 -- build intelligence, never a fixed ramp. The cheapest open rung on this board. Adoption: `LIB-WIRED importers=5 nonval=2` — fully adopted (top of its ladder).	●	◉	◉	◉	●
full PBR map set and normal-map lighting Measured: `ntb_bake_pbr` exists in `runtime/nx_nxa_texbake_lib.nx`, verified at emit. R4 gap watching ntb_bake_pbr (residue b and d). Albedo alone reads as clay; the measured floor for a complete skin is 4 maps across 4 regions at 4096px, and 4096 is untested since resolution is already a knob. Rivals ship fixed files, so the exceed axis available to us is regeneration PER GENOME at load. Adoption: `LIB-WIRED importers=5 nonval=2` — fully adopted (top of its ladder).	●	◉	◉	◉	●
atlas sampling in the world NPC shader Measured: `cw_npc_atlas` exists in `runtime/nx_craft_emit.nx`, verified at emit. R6 gap watching cw_npc_atlas (residue c). The CORE renderer samples the atlas; the craft-world NPC wasm shader does not yet -- so textured skin is proven in the lab and absent from the shipped world. A class jump that lands in one surface and not its siblings just moves the complaint one URL over. Adoption: `LIVE` — fully adopted (top of its ladder).	●	◉	◉	◉	●
subsurface scattering skin shader Measured: `sr_sss_preint` exists in `runtime/nx_sdfrender.nx`, verified at emit. R5 gap watching sr_sss_preint. Per-channel wrap diffuse exists in the SDF lane already; pre-integrated SSS bound to the baked normal map is the rung that stops the clay read. Adoption: `LIB-WIRED importers=46 nonval=25` — fully adopted (top of its ladder).	●	◉	◉	●	●
BODY
soft-body dynamics bound to skin vertices Open — watching `runtime/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. R7 gap watching dyna_bind_skin. nx_nxa_dyna already WRITES a per-bone DYNA section into the asset, and its promoted binary sits at 23192B behind its 28747B build -- a stale promote to refresh. The solver is gate-proven; binding it to skin geometry is the open rung. SHIPPED 2026-08-23 but NOT under this contract symbol: the capability lives in organ nx_dyna_bind_skin (live 6aa76da6, gate nx_dyna_bind_skin_gate 15/15 GREEN) which embeds skin vertices in the XPBD cage [xpbd] and measures the plant by ringdown. At the linear limit LARGE_SOFT reads fn 4137/4067 mHz inside tissue_fn_mhz 4000-5200 and zeta 480/488 inside tissue_zeta_permil 350-600, both branches, nothing tuned. THIS ROW CANNOT FLIP AS WRITTEN because it watches symbol dyna_bind_skin inside nx_nxa_dyna.nx and the shipped symbol is elsewhere; retargeting the contract is the board owner call, and a stub carrying the contract name would light the cell green with no capability behind it. RESEARCH BASE ADMITTED 2026-08-25, every item fetched through the sovereign fetcher, content-pinned and journalled at fetch time: PhysRig [physrig] is the differentiable skeleton-embedded-in-a-tetrahedral-volume method, and it fits Young's modulus and Poisson's ratio but NEVER damping, so our ringdown measures an axis it does not state; SoftSMPL [softsmpl] is the learned nonlinear soft-tissue regressor for parametric humans whose own abstract names dataset scarcity and overfitting as its weakness, which is the weakness a derived plant does not have; the three-dimensional human breast-tissue study [breasttissue] is subject truth that CONFIRMS incompressibility and non-negligible anisotropy while QUALIFYING our sign-asymmetric branch, since it found a bimodular model unnecessary where cai2018 motivates one, and that disagreement is recorded for adjudication rather than resolved silently; the Blender spring-bones addon [springbones] is the artist-knob game convention the derived plant departs from; and an independent methodology survey [daeresearch] states plainly that real-time soft-body organ simulation is not yet possible, which is the context that makes our in-process reduced-order claim honest. watching dyna_bind_skin	○	◉	◉	●	●
cloth and hair on the same solver Open — watching `runtime/nx_nxa_dyna.nx` : `dyna_cloth`, re-measured on every compare beat. Ship that symbol and this mark flips itself; the `comparewatch-` plane row flips with it. R8 gap watching dyna_cloth -- wardrobe layers and hair strands driven by the same dynamics, never a second system. watching dyna_cloth	○	◉	◉	●	●
ANIM
skeletal animation retargeted from real mocap Measured: `m_joints` exists in `runtime/nx_bvhmotion.nx`, verified at emit. BVH and VMD decoded sovereignly and retargeted at 30fps. Parity on presence; rivals ship authored clip libraries instead. 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). adoption REGISTERED-DARK	●	●	●	●	●
paired two-character interaction Open — watching `runtime/nx_anim_paired.nx` : `ap_paired_clip`, re-measured on every compare beat. Ship that symbol and this mark flips itself; the `comparewatch-` plane row flips with it. R3 gap watching ap_paired_clip. The defining capability of all four titles and the largest true greenfield here: two skeletons, contact constraints, no interpenetration. The referee already exists -- nx_clipcheck is LIVE and is the ruler that grades it. watching ap_paired_clip	○	◉	◉	◉	◉
facial expression morphs Measured: `fa_expr_morph` exists in `runtime/nx_faceanat.nx`, verified at emit. R9 gap watching fa_expr_morph. Identity morphs already exist as anthropometric axes; FACS-style expression axes do not. Adoption: `LIB-WIRED importers=12 nonval=2` — fully adopted (top of its ladder).	●	◉	◉	●	●
SIM
genome breeding with proven inheritance Measured: `gn_breed` exists in `runtime/_hdl_build/nx_breeding.nx`, verified at emit. gn_breed is seeded per-trait inheritance with bounded mutation, gated 8 of 8 -- children measurably closer to parents than to wild genomes, zero duplicate offspring over 100 seeds, 25 chained generations in bounds. The advantage over a finite authored roster stays a HYPOTHESIS until a vsbench cell measures it. Adoption: `LIB-GATE-ONLY importers=1` — PARTIAL: imported only by validation organs (gates, tests, benches): wire it into a shipping program. adoption LIB-GATE-ONLY importers=1	●	○	○	◉	●
transformation as a genome delta Measured: `gn_morph` exists in `runtime/_hdl_build/nx_breeding.nx`, verified at emit. gn_morph moves ONE trait and leaves every other word bit-identical, so a transformation re-emits through the same body rules instead of shipping a second authored asset set. Adoption: `LIB-GATE-ONLY importers=1` — PARTIAL: imported only by validation organs (gates, tests, benches): wire it into a shipping program. adoption LIB-GATE-ONLY importers=1	●	●	●	●	○
creature battle with a real type chart Measured: `cb_eff` exists in `runtime/_hdl_build/nx_creature_battle.nx`, verified at emit. cb_eff is the effectiveness matrix proven in the M2D slice. Breeders and Tropical Monster Girls both carry combat; the two Illusion titles do not. Adoption: `LIB-WIRED importers=6 nonval=4` — fully adopted (top of its ladder).	●	○	○	●	●
relationship and daily-schedule NPCs Open — watching `runtime/nx_relsim.nx` : `rs_schedule_tick`, re-measured on every compare beat. Ship that symbol and this mark flips itself; the `comparewatch-` plane row flips with it. R10 gap watching rs_schedule_tick. Relationship state exists in nx_charmem and a conserved economy exists in nx_worldsim, but nothing yet drives NPCs through a day arc. watching rs_schedule_tick	○	●	●	●	◉
relationship memory that persists Measured: `cm_ctx` exists in `runtime/nx_charmem.nx`, verified at emit. cm_ctx folds preferences and remembered events into a context block that survives sessions. Built and registered; no shipped game consumes it yet. 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). adoption REGISTERED-DARK	●	●	●	○	●
freeform LLM NPC dialogue Open — watching `runtime/nx_charmem.nx` : `cm_npc_turn`, re-measured on every compare beat. Ship that symbol and this mark flips itself; the `comparewatch-` plane row flips with it. R11 gap watching cm_npc_turn. None of the four RIVAL COLUMNS contests this axis -- Koikatsu, Honey Select 2, Breeders and Tropical Monster Girls all ship authored dialogue trees, and that narrow reading stands. CORRECTED 2026-08-22, and the correction is published rather than quietly repaired: this note used to generalise that to the clearest uncontested lane in the field, and THAT IS REFUTED. SecretWaifu ships a real-time AI companion whose entire product IS live conversation with an expressive character, witnessed 7 occurrences of 7 in a pinned mirror including the visible hero paragraph [swaifu]. The four columns are a character-maker and breeding-sim field which structurally EXCLUDES AI-companion competitors, so this axis is uncontested only WITHIN these columns and is contested in the wider field. The uncontested-lane rank bonus R11 draws was therefore computed over the wrong population -- a narrower subject than the word field implies. watching cm_npc_turn	○	○	○	○	○
EMERGE
agent memory stream with ranked recall Measured: `em_recall` exists in `runtime/nx_elara_memory.nx`, verified at emit. em_recall ranks a session's memory plane by term overlap plus recency over a leak-free segment scan. The Smallville-class memory stream [generative-agents23] EXISTS; its own header names semantic embedding as rung 2, and a sibling lane has since shipped trained dense vectors, so that rung is unblocked rather than unstarted. Not one title in this field ships persistent agent memory at all. 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). adoption REGISTERED-DARK	●	○	○	○	○
semantic recall over the memory stream Open — watching `runtime/nx_elara_memory.nx` : `em_recall_dense`, re-measured on every compare beat. Ship that symbol and this mark flips itself; the `comparewatch-` plane row flips with it. R14 gap watching em_recall_dense -- swap the lexical ranker for the trained dense model already live in the search lane (maxsim over a 102318-term vocabulary). A scorer swap composing an existing organ, never a second implementation. watching em_recall_dense	○	○	○	○	○
observation engine turning world events into percepts Open — watching `runtime/nx_obsengine.nx` : `obs_tick`, re-measured on every compare beat. Ship that symbol and this mark flips itself; the `comparewatch-` plane row flips with it. R15 gap watching obs_tick -- the missing front of the loop. A procedural world event has to become something an agent can perceive and write down. nx_worldsim and nx_worldpipe already produce the events; nothing yet renders them as observations [emergent-tool-use19]. watching obs_tick	○	○	○	○	○
reflection loop over an agent's own memories Open — watching `runtime/nx_dr_refute.nx` : `dr_reflect_agent`, re-measured on every compare beat. Ship that symbol and this mark flips itself; the `comparewatch-` plane row flips with it. R16 gap watching dr_reflect_agent. The reflection PRIMITIVE already ships: nx_dr_refute runs N-verifier adversarial voting with a decorrelation discount so correlated refutations count less, a Co-Scientist-class kill stage. What is missing is binding it to an NPC's own memory stream so the agent revises its beliefs, not only ours [generative-agents23]. watching dr_reflect_agent	○	○	○	○	○
planner emitting structurally-winnable story graphs Measured: `pg_gen` exists in `runtime/_hdl_build/nx_plotgen.nx`, verified at emit. pg_gen lays a main spine whose advance is always ungated, and permits a branch at node k to require only flags raised strictly before k, so an unwinnable quest is STRUCTURALLY impossible -- proven by walking 200 of 200 seeds to an ending rather than by inspection. Rivals author graphs by hand and QA them afterwards. Adoption: `LIB-WIRED importers=3 nonval=1` — fully adopted (top of its ladder).	●	●	●	●	●
oracle translating agent intent into engine actions Open — watching `runtime/nx_npc_oracle.nx` : `no_intent_exec`, re-measured on every compare beat. Ship that symbol and this mark flips itself; the `comparewatch-` plane row flips with it. R17 gap watching no_intent_exec -- the narrative-translator layer. An agent saying it will cook a meal must become coordinate-level pathfinding and object interaction; nx_pathfind and nx_inventory are the certified parts it would drive [emergent-tool-use19]. watching no_intent_exec	○	○	○	○	○
graph-state evaluator derived from real play logs Open — watching `runtime/nx_story_vm.nx` : `sv_arc_derive`, re-measured on every compare beat. Ship that symbol and this mark flips itself; the `comparewatch-` plane row flips with it. R18 gap watching sv_arc_derive. nx_story_vm EXECUTES authored graphs; deriving a graph FROM logged play is what detects that a player's unscripted action birthed a genuinely new arc. The estate already emits mermaid graphs elsewhere, so this is composition rather than construction. watching sv_arc_derive	○	○	○	○	○
long-horizon drift and safety-erosion harness Open — watching `runtime/nx_emergebench.nx` : `eb_longrun`, re-measured on every compare beat. Ship that symbol and this mark flips itself; the `comparewatch-` plane row flips with it. R19 gap watching eb_longrun. The published finding it must reproduce is that explicit prohibitions given to agents DEGRADE over long horizons under perceived structural failure [emergence-world]. Ours is the sharpest available test case precisely because the adult invariant is not a prompt prohibition an agent can reason around -- it fails closed at the compiler. Bit-exact deterministic replay also makes any divergence bisectable, which the reference platforms cannot do. watching eb_longrun	○	○	○	○	○
hidden-role social-deduction arena Open — watching `runtime/nx_emergebench.nx` : `eb_werewolf`, re-measured on every compare beat. Ship that symbol and this mark flips itself; the `comparewatch-` plane row flips with it. R20 gap watching eb_werewolf. A grade-1 EXTERNAL referee for negotiation and strategic deception, run on someone else's published protocol rather than our own scoring [game-arena-werewolf]. Never invent a ruler. watching eb_werewolf	○	○	○	○	○
WORLD
procedural world with real lighting Measured: `wp_shade` exists in `runtime/nx_worldpipe.nx`, verified at emit. wp_shade is sun-facing diffuse off the real heightfield, unity-centred and gate-proven in BOTH directions (sunward brightens, shadeward darkens -- the non-gameable form). Breeders ships an open world; the Illusion titles are room-scale by design. Adoption: `LIB-WIRED importers=8 nonval=3` — fully adopted (top of its ladder).	●	○	○	◉	●
characters drawn in-world by the sovereign renderer Measured: `mob_draw` exists in `runtime/nx_wasm_craft.nx`, verified at emit. mob_draw projects every NPC into the CPU raycaster's own framebuffer with per-mob ray occlusion -- no GPU and no third-party layer in the path. Gate-measured rather than asserted: T59 renders the same frame twice differing ONLY in whether the girls are near, and counts 621 pixels of difference, so the tooth fails when they are absent from the PICTURE rather than merely present in state. That distinction is the whole point -- T16, T17, T28 and T29 all verify mobs in state and stayed green through a regression that put every girl outside the written frame. Adoption: `LIB-WIRED importers=7 nonval=3` — fully adopted (top of its ladder).	●	◉	◉	◉	●
skinned mesh characters in-world, not billboards Measured: `wc_mob_mesh` exists in `runtime/nx_wasm_craft.nx`, verified at emit. ÃƒÂ¢Ã‹Å“Ã¢â‚¬Â¦R21 SHIPPED 2026-08-15 -- wc_mob_mesh is REAL GEOMETRY, and the flip is the receipt while the gate is the proof: an 11-bone humanoid rig built on nx_skeleton's LBS (the part prepared for exactly this: base+offset, no sys_mmap, wasm-safe), 133 ring vertices whose proportions all derive from ONE height (RIG_H=240, the billboard's world volume, kept so every collision/door/schedule tooth stays proven) and the 7.5-head canon unit RIG_HU=RIG_H2/15; an integer edge-function rasterizer interpolating depth per pixel against the terrain depth the voxel pass banks in the framebuffer's high bits (the zero-byte depth plane shipped the same day), so limbs occlude PER PIXEL and mob-vs-mob ordering needs no sorting; gait and arm counter-swing on the same phase that drives the shipped bob, bands cited (hip 30deg=IT_PI/6, arms half); genome shading unchanged in mapping. Gate 63/63 GREEN with the teeth that make the claim falsifiable: T63 mesh-not-billboard fill_permil=545 of her own bbox (a quad is ~1000, specks <=100 -- BOTH bounds bite), T61 genome-distinct 196px with same-girl-twice=0, T62 pins the RIG_BYTES literal to sk_bytes() so the one deliberate duplicate can never drift, T60 headroom 1,599,808. The billboard was DELETED, not parked behind a dead conditional. Shipped to all five worlds, wasm-vm gate 5/5 on the loaded module. ÃƒÂ¢Ã…Â¡Ã‚Â HONEST MEASUREMENT RESULTS, BOTH DIRECTIONS: the frame-level referee is INSENSITIVE to this rung -- character-framed CHARJUDGE stayed 46 (palette 296->312, contour 866->850) because composition is a FRAME statistic and a 180px figure cannot move it in a 36,000px voxel scene; the figure-quality instrument is the gate's shape teeth -- AND THE SUBJECT-CROPPED JUDGE MODE SHIPPED THE SAME DAY: nx_world_snap char now renders at q=1 NATIVELY (1200x750 -- the browser's adaptive controller trades rays for frame time, but an offline instrument runs once and can afford every ray), banks the framed render, parks the girl, re-renders the identical scene, and every differing pixel IS her -- the crop rectangle is MEASURED, never guessed, with a margin of exactly one judge composition cell (16px) so the crop is derived from its consumer's own geometry. First run: pixels=3571 (20x the q=5 subject data), bbox 55x152, fill_permil=427 -- corroborating gate T63's 545 from a SECOND independent instrument -- crop 87x184 written as a derived sibling (<out>.crop.png), a framed-but-zero-diff run announces pixels=0 and writes NO crop, and the world path was proven byte-identical by behaveprobe before promote. FIRST FIGURE-QUALITY BASELINE: crop CHARJUDGE 0 (composition 0 -- the flat-shaded figure's limbs are narrower than the judge's 16px cells, so nothing reads smooth -- palette 195, contour 333), against the lab head's 289 and the corpus mean 138. That zero is the honest SOTA gap AS A NUMBER: the in-world girl's beauty is now measured, the chain edit -> gate -> snap -> crop -> judge is closed end to end, and the roadmap rungs (surface shading, SSS, mesh detail, a closer camera) finally have a ruler that only moves when SHE improves. ÃƒÂ¢Ã‹Å“Ã¢â‚¬Â¦FIRST RUNG UP THE LADDER SHIPPED 2026-08-16 -- PER-TRIANGLE LAMBERT, LIT BY THE WORLD'S OWN SUN: face normals from the skinned verts (translation-invariant, so rig space suffices), oriented outward against the segment centroid (winding-proof: a sign test against the radial vector is correct for every segment), shaded against the SAME SUX/SUY/SUZ vector wshade2 lights the terrain with -- characters sit in the world's light, no second private sun and no new light constants. The magnitude uses the alpha-max-beta-min approximation with its error STATED (a few percent, feeding only the brightness factor), chosen over adding yet another hand-rolled isqrt beside the several the estate already carries scattered (nx_pm_isqrt, df_isqrt -- hoisting a shared it_isqrt is the named follow-on). What CANNOT be tuned or drifted is pinned by gate T64 in wp_shade's own non-gameable form: sun-facing strictly brighter per channel AND the away side EXACTLY equal to the ambient term -- an equality no approximation error can fake. Gate 64/64, shipped to all five worlds. MEASURED ON THE SAME FRAMING, judged same-day: crop palette 195->208, contour 333->733 (coherent 20->44 permil, edges 2463->4077 -- the tonal structure is real), composition an HONEST ZERO STILL, with the cause diagnosed rather than hand-waved: smooth cells need flat-or-gently-shaded regions wider than the judge's 16px cell, and at the capture's four-block distance her torso spans barely one cell. The next lever is the capture's portrait distance -- an INSTRUMENT calibration, deliberately not bundled into the same measurement as a renderer change, declared here before it moves so the ruler's history stays readable. ÃƒÂ¢Ã‹Å“Ã¢â‚¬Â¦AND THEN IT MOVED, AND THE BAR FELL, 2026-08-16: the portrait distance was DERIVED (torso must span >= 2 judge cells: 44 q8 at foc 600 needs vz <= 825; two blocks gives a 51px torso, ~3 cells) and the world path proven byte-identical before promote. Portrait capture: pixels=14034 (4x the old framing, as the arithmetic predicted), fill_permil=412 -- the THIRD independent framing to land the shape statistic in the same band (427, 545, 412), a free stability cross-check. FIGURE-CROP CHARJUDGE 0 -> 251* (composition 673 with smooth 101 permil -- her lambert-shaded regions finally span whole cells -- contour 1000, palette 251), EXCEEDING THE PRE-DECLARED FLIP BAR OF 138. The rule was declared before the experiment and the calibration before the measurement; honoring it is the same discipline as refusing the premature symbol-keyed flip the day before. GPE_EMIT_WEBGL FLIPPED TO 0 BY ITS OWN CONTRACT: all five worlds now ship SOVEREIGN-ONLY (~305KB -> ~230KB, the GLSL retired), the footer rewritten so no control lies (a ?gl=1 link on a page with no GLSL would be a switch that silently does nothing), and nx_sovworld_gate closed its standing RED honestly -- 5/5 GREEN, wc_mob_mesh present=1, page carries WebGL=0. THE DEFAULT VIEW OF EVERY WORLD IS NOW THE NISHILANG-DRAWN, SUN-LIT, GENOME-DISTINCT SKINNED GIRLS -- what the operator asked to SEE is what every visitor sees, and if the lived result disappoints, the flip is one flag back with every justifying number on this board. The retired WebGL NXA character renderer (dual-quat, eyeballs) remains in the tree as the reference to beat on richness -- eyes, hands, hair as geometry are the named rungs the ruler will price next. ÃƒÂ¢Ã‹Å“Ã¢â‚¬Â¦EYES SHIPPED 2026-08-16 -- the ruler priced its first richness rung and caught a SILENT REGRESSION in the same stroke: the iris verts the mesh had carried since it landed sat INSIDE the radius-16 head sphere, so eyes NEVER rasterized -- a missing feature with no failure mode, invisible to every state tooth. Verts moved ONTO the surface ((RIG_HU/2, RIG_HU3/8, +-RIG_HU9/32)=(16,12,+-9), derived from the same one-height canon), 8 sclera verts added, sclera quads drawn UNLIT (skin mixed 700/1000 toward near-white -- sclera reads bright regardless of sun angle) and iris marks drawn after. Gate now 65/65: NEW TOOTH T65 eyes-in-picture pins it forever -- iris-colored pixels inside girl-A's bbox must EXCEED the girl-less frame's count (girl=4, girl-less=0 at gate res), so the silent-loss class that hid this cannot recur. Shipped to all five worlds, wasm-vm gate 5/5, read-back verified; nx_world_snap rebuilt so the instrument carries the same engine. JUDGED ON THE UNCHANGED PORTRAIT CALIBRATION: crop CHARJUDGE 251 -> 253 (palette 251->253 is the MIN axis and moved by the new sclera/iris tones; composition 673->633, contour 1000, fill_permil 412 stable). Honest read: +2 is the referee pricing a handful of iris pixels at portrait distance as barely visible -- this rung's value is the REGRESSION FIX plus T65's pin, not a beauty jump; face_axis stays 0 WITH ATTRIBUTION (pyramid_levels 3, candidates 1, min_level_w 35): the probe FIRES but one candidate cannot pair -- the documented contrast limit, not a silent zero. Next rungs unchanged: hands, hair-as-geometry, SSS (sr_sss_preint), closer portrait framing. ÃƒÂ¢Ã‹Å“Ã¢â‚¬Â¦SKIN WRAP-SSS SHIPPED 2026-08-16 (same day): per-channel WRAP DIFFUSE on the SKIN segments only (head, shins, forearms -- rig_seg_reg 0), the zero-curvature floor of pre-integrated subsurface scattering (Penner 2011: diffusion widens the terminator per channel, red furthest), as named consts because the wasm arena cannot read conf at runtime -- compile-time IS that module's data lane. wc_lambert itself UNTOUCHED (T64's subject stays pinned); outfit and hair keep it. NEW TOOTH T66 pins the three consequences no tuning can game, on one grey input: fully-away EQUALS plain lambert BIT-EXACTLY (a red tint dies), the terminator is strictly brighter per channel with red gaining STRICTLY more than blue (a uniform brightener dies; measured gain r=43 vs b=16, and the live gate printed exactly the hand-derived integers), and brightness is monotone in sun angle. Gate 66/66, all five worlds shipped (VM gate 5/5, read-back verified), snap twin rebuilt. JUDGED, UNCHANGED CALIBRATION: crop CHARJUDGE 253 -> 257 (palette 253->257 IS the MIN axis and the wrap is exactly a palette mechanism -- buckets 130->132; composition 593 and contour 1000 ride above the MIN). The ladder the ruler has priced: 0 -> 251 (lambert+portrait) -> 253 (eyes) -> 257 (wrap-SSS), bar 138. ÃƒÂ¢Ã…Â¡Ã‚Â The sr_sss_preint WATCH ROW STAYS OPEN on purpose: it names the SDF lab lane's pre-integrated shader bound to a baked normal map -- this craft rung is the same physics at the voxel engine's floor, not that contract's symbol, and lighting a watch green from a different organ would be the name-keyed completion defect this board already refused once. ÃƒÂ¢Ã‹Å“Ã¢â‚¬Â¦ÃƒÂ¢Ã‹Å“Ã¢â‚¬Â¦WOMEN, NOT GUMBY -- 2026-08-16, operator eyeball as referee ("wtf everything is broken with green gumby vs women"), and the portrait AGREED: world-palette clay, arms floating in daylight, a straight tube torso. FOUR mechanical defects, each now TOOTHED: (1) SKIN WAS STRUCTURALLY IMPOSSIBLE -- c_skin mixed the KIND's world spec tones (wood-brown here, teal/ember/violet elsewhere; human skin occurs in NO world's palette). Fixed by COMPOSING the estate's own unwired incumbent: nx_skin_ita (charsim R0, published ITA->Lab->sRGB standard, gate-proven) -- genome skin bits now span DARK->VERY-LIGHT ITA classes with the lib's own face/limb offsets; the world keeps its identity in the OUTFIT. T69 pins L strictly rising (31.4->74.5), classes strictly distinct, r>g>b at both ends. (2) LIMBS FLOATED: SHOULDER_Z 40 vs chest 22, HIP_Z 26 vs hip 20-10. Now DERIVED (deltoid embed: arm center ON the taper surface; thighs inside the pelvis footprint) and pinned by T70 -- WITH A SUBJECT-MATCH LESSON PAID IN FULL: the first fix held the width inequality at the shoulder line while the bones attached at the HEAD JOINT beside the r10 neck, and the tooth was GREEN while the picture stayed broken. A radius sum proven at a height the arms never touch is an instrument whose subject does not match its claim; RIG_ARM_Y==RIG_SHBASE_Y is now a conjunct. (3) TUBE TORSO -> HOURGLASS: per-END ring radii + stacked segments (hip 20 -> waist 14 -> bust 20 -> shoulder 18, neck RIG_HU/3 by Loomis), waist/hip = 0.7 EXACT (published female mean; integer arithmetic lands it as an EQUALITY, pinned by T71). (4) HAIR: natural melanin span (near-black->blonde, 8 genome hue stops) replacing world-tone-toward-black, PLUS the documented-but-discarded LENGTH bits 3-4 now drive geometry -- back fall hugging the body + front side-locks onto the chest (tops on the head bone, bottoms on the CHEST bone so a head turn cannot drag chest-lying hair). T68 proves length reaches the PICTURE: 319 pixels short-vs-long at q=1 native, neg-control 0 -- and its own build history is the measurement lesson of the day: the first curtain read 0 px because it was DEPTH-CLIPPED behind the bust (the tooth fired exactly as designed), and the locks read 0 px at the gate's q=5 because a 4.6 q8 strip is HALF A PIXEL there (proven by a compiled-in projection diagnostic, both edges on screen x=121) -- the fix was rendering the tooth at q=1, THE RESOLUTION THE BROWSER SHIPS, never fattening geometry to please a coarse ruler. T67 pins vertex capacity as a derived equality (built==RIG_VERT_TOTAL==PROJ_MAXV==211): the silent projection-overflow class (the pmboard zero-byte lesson) closed BEFORE it ever fired. Gate 71/71 GREEN, all five worlds shipped (~241.5KB, VM gate 5/5, read-back verified), snap twin rebuilt. JUDGED HONESTLY, BLINDNESS NAMED: figure-crop CHARJUDGE reads 234 vs 257 pre-rework (composition 593->1000 -- the coherent figure finally fills whole judge cells -- palette dipped as the MIN because the tighter 89px crop carries fewer buckets 132->120, a crop-size sensitivity of that axis). THE REFEREE IS STRUCTURALLY HUE- AND SEMANTICS-BLIND: green clay and a woman score alike on palette/contour/composition, WHICH IS HOW GUMBY SHIPPED THROUGH A GREEN BOARD. The rulers that DO see this rung are the new teeth (T68-T71) and the served portrait itself; face_axis attribution improved 1->2 candidates (both eyes register). NEXT: head/face shape (the box reads worst now), bust geometry (circular rings cannot carry it), Kovac skin-pixel picture tooth (nx_color_kovac_skin_pixel is the found incumbent), judge crop-size normalization. ÃƒÂ¢Ã…Â¡Ã‚Â CONSEQUENCES ARMED, THEN ADJUDICATED AGAINST MY OWN FIRST READING: nx_sovworld_gate went RED 4/5 by its own design (T5: the allowance retires when wc_mob_mesh ships) and I first published that the GPU path draws terrain only -- WRONG, refuted by the emitter's own header, which records that the WebGL path is the ONLY code drawing the rich skinned NXA characters (dual-quaternion skinning, eyeball geometry, per-genome skin/outfit/hair at canvas resolution) and that deleting it before parity is the regression the operator reported THREE times on 2026-08-14. So the two contracts disagree -- the gate keys on the SYMBOL existing, the emitter keys on PARITY -- and the newer, lived one controls: a name-keyed completion signal must not delete an incumbent. GPE_EMIT_WEBGL stays 1; T5 stays RED as the standing parity work-order, not silenced; and the flip condition is PRE-DECLARED and published rather than tasted: the emitter flips to sovereign-only when the sovereign figure-crop CHARJUDGE meets or beats the character corpus mean on this board (138 at declaration). Meanwhile the operator's real complaint -- none of this session's work visible on the default view -- is fixed the non-regressive way: every world page now carries a visible one-click switch ("See it yourself: sovereign render / GPU render") instead of a URL secret, shipped to all five worlds and verified in the served bytes. ÃƒÂ¢Ã‹Å“Ã¢â‚¬Â¦ÃƒÂ¢Ã‹Å“Ã¢â‚¬Â¦DELIVERY-LAYER FIXES 2026-08-17 (operator: "the world /beach and all that are incredibly pixelated ... we have a 5080"): THREE stacked causes, each measured before touched. (1) THE ADAPTIVE CONTROLLER PARKED EVERY MACHINE AT 41 PERCENT OF ITS OWN BUDGET: it coarsened above 24ms but refined only below a magic 11ms line, so any device between them sat pixelated forever. REPLACED with the renderer's own cost model (rays scale 1/q^2: refine when FMAq^2/(q-1)^2 fits under the budget line minus one 4ms margin) -- and the first attempt was WRONG in a way gate T20 caught (2/3 RED): the coarsen branch's FMA=256 reset is a SEEDED value, not a measurement, and same-call refine read it and undid every coarsen step at q>3. Refine now never acts on the call that coarsened; T20 3/3, gate 71/71. (2) THE PAGE FORCED image-rendering:pixelated -- nearest-neighbor blocks by explicit CSS; removed, browser bilinear does the upscale. ÃƒÂ¢Ã…Â¡Ã‚Â THE FIX DID NOT SHIP ON THE FIRST EMIT because the page CSS is compiled INTO the promoted nx_craft_emit binary while the wasm is compiled from source at emit time -- an emit after an emitter-source edit ships new ENGINE and old PAGE silently; the emitter itself must rebuild first. (3) EVERYTHING SHIPPED TODAY WAS INVISIBLE TO THE OPERATOR ANYWAY: sites.elf served OLD in-process page content that survived every atomic file replacement (the deleted-inode class -- the daemon's copy outlives the rename), proven by a never-seen probe path serving old bytes at +990B over any disk file; one health-checked /api/restart service=sites cleared it, verified fresh from BOTH vantages afterward, all 16 services UP. The +990B wrapper itself adjudicated BENIGN: a 940x150 telemetry trace strip (gait series drawn as line graphs) injected by earlier instrumentation -- pixelated is CORRECT for a trace plot and it is not the game canvas. STILL OPEN AND NAMED: a CPU raycaster cannot reach native 1200x750 in 24ms on typical hardware -- the ceiling-breaker is the WebGPU-compute rung (OUR sovereign raycaster emitted as compute shader, the 5080 executing our code), now the delivery half of the R25 lane. ÃƒÂ¢Ã…Â¡Ã‚Â Also noted: the emitter's wasmfit line still recites the framebuffer-tail total (10,790,528) and does not count the rig region; the binding bound is gate T60 over CRAFT_TOTAL=10,983,104. Old note follows for the parts still open (nx_trimesh tm_bind full-mesh rung, motion-facing yaw). ÃƒÂ¢Ã‹Å“Ã¢â‚¬Â¦GENOME-DISTINCT GIRLS SHIPPED 2026-08-15, and the defect it fixed was that the variety already EXISTED in the data and was discarded at the last step: MC_GENE is a documented 20-bit appearance genome (bits 0-2 hair hue, 7-8 outfit tint, 9-10 skin tone) bred through gn_breed, T29 measures genomes distinct=12/12, and mob_draw painted mob_col(kind, half) -- two flat tones per KIND -- so twelve genetically distinct girls rendered as three identical pairs. Three region colours are now derived per girl from her own genome, computed once per mob and not per pixel, introducing NO new palette constants: the kind's own two spec tones stay the ramp and the genome only chooses where on it each region sits, so world identity still owns its colours while the individual owns her variation. Regions are anchored on geometry the module already used -- the eye line at -hh2/2 and the pre-existing yy>0 split -- so no body fractions were invented. T61 proves it end to end by placing two SAME-KIND, DIFFERENT-GENOME girls at the identical spot: same-kind-diff-genome-pixels=489 with neg-control-same-girl-twice=0, so the difference is attributable to the genome and not to render noise; under the old renderer that first number is 0 by construction. Gate 61/61 GREEN, wasm-vm gate 5/5 on the SHIPPED module, headroom unchanged at 1792384. ÃƒÂ¢Ã…Â¡Ã‚Â MEASUREMENT GAP FOUND IN THE SAME BREATH: nx_world_snap -> nx_charjudge scores the before and after frames IDENTICALLY in every field (composition 246, palette 343, contour 933, palette_buckets 176, edge_total 10761) because the world snapshot frames no girl at all -- the capture instrument cannot see character work, which is the same blind spot T59's own comment names for the colour-variety teeth. The referee is only as good as the frame it is given; a CHARACTER-FRAMED capture is owed, and the mechanism already exists inside T59, which places a girl four blocks ahead at eye height. mob_draw is still a billboard; all four rivals render skinned meshes, and this estate's own WebGL path did too until it was retired for sovereignty. The parts are already sovereign and arena-native: nx_skeleton is 3D linear-blend skinning (32 bones, 2048 verts, 188736 bytes per rig, base+offset addressing with NO sys_mmap) and nx_trimesh now binds its vertex arrays to caller memory through tm_bind, which is what removed the two measured blockers -- a 136 MiB fixed capacity and a sys_mmap that wasm does not have at all. What remains is the compositing rung: seed a depth buffer from the voxel pass so terrain occludes limbs correctly. Adoption:* `LIB-WIRED importers=7 nonval=3` — fully adopted (top of its ladder).	●	◉	◉	◉	●
PLATFORM
one loop from keyboard, gamepad, touch and VR Measured exceed: `ia_bind` in `runtime/nx_input_abstract.nx`, verified at emit. ia_bind turns a raw device code into a semantic action; all four device classes drive the same loop to a bit-identical checksum, mutation-proven. Rivals fork the loop per platform. Adoption: `LIB-WIRED importers=6 nonval=3` — fully adopted (top of its ladder).	◉	●	●	●	●
VR headset runtime Open — watching `runtime/nx_sdfrender.nx` : `sr_vr_present`, re-measured on every compare beat. Ship that symbol and this mark flips itself; the `comparewatch-` plane row flips with it. R12 gap watching sr_vr_present. Geometrically correct off-axis stereo pairs are proven on our own renderer; a headset present-loop is not. Honey Select 2 ships VR and a headset is available for real-device rungs. watching sr_vr_present	○	○	◉	○	○
TRUST
adult invariant fails closed at the compiler Measured exceed: `ws_coherent` in `runtime/_hdl_build/nx_wardrobe_state.nx`, verified at emit. ws_coherent refuses a self-contradicting state and the setter refuses exposing a gated zone under age 18 with the state left bit-identical -- mutation-proven, and it holds through a whole game loop. Rivals hold this by policy; ours holds by proof. Adoption: `LIB-WIRED importers=3 nonval=2` — fully adopted (top of its ladder).	◉	●	●	●	●
a character is a seed, not an asset pack Measured exceed: `gs_save` in `runtime/_hdl_build/nx_gamesave.nx`, verified at emit. gs_save round-trips bit-exact with a caller-supplied timestamp, so there is no hidden clock and a shared string reproduces the being exactly. VaM-class scene sharing moves gigabytes to do the same job. Adoption: `LIB-WIRED importers=28 nonval=8` — fully adopted (top of its ladder).	◉	●	●	○	○
mechanical character-image judge Measured: `cjc_hw` exists in `runtime/_hdl_build/nx_charjudge.nx`, verified at emit. cjc_hw scores character images mechanically instead of by eyeball. WIRED AND CORRECTED 2026-08-15, and wiring it is what exposed it: run as a referee over the real render corpus it was INVERTED on this estate's own shipped work. The contour axis took 1px finite differences on RAW luma, so PBR surface texture broke the orientation-stable runs -- two independent untextured/textured A/B pairs of the SAME body scored 143->16 and 223->33, i.e. the judge graded the landed skin bake an 8.9x and 6.8x REGRESSION while composition, palette and detail all said it improved. Fixed by estimating the gradient on a smoothed channel, radius DERIVED from run_min so the two can never drift: the same pairs now read 143->173 and 223->246, correctly ordered, with the oracle control unmoved at 298 and ours closing 116->289. The FACE axis is demoted out of the headline MIN and now only reports: it scored 1000 on a voxel landscape with no character in frame and 0 on a face close-up, and three successive probe geometries each traded one false negative for another -- an uncalibrated classifier must report numbers, never verdicts, so re-arming it needs a labelled corpus and a measured FP/FN rate first. nx_charjudge_gate is now base-class (gv_check/gv_bite/gv_verdict), 11/11 GREEN, with teeth at a SECOND resolution and a bite-proven negative control -- all 8 original teeth ran only at 256px == ref_w, which is exactly the blind spot that let both defects ship green. R0/R1/R4 are now gradeable on composition x palette x contour. BASELINE MEASURED, FULL POPULATION: nx_charjudge_census forks the PROMOTED judge once per image over every .png in the world docroot (selection rule stated in knowledge/charjudge_corpus.list -- not curated, not sampled, non-character renders deliberately left in) and returns corpus=169 judged=169 failed=0 truncated=0, partition RECONCILES, headline min=0 max=794 mean=141. ÃƒÂ¢Ã…Â¡Ã‚Â CORRECTED TO mean=140 ON 2026-08-15, AND THE REASON MATTERS MORE THAN THE POINT: nx_world_snap read its frame extent from ww()/hh(), which are the wasm export surface and are hardwired to ARENA 0, while every render went into a fresh arena the organ never registered as 0. The two agreed by luck. It reported w=480 h=300 when the arena it actually rendered is 240x150 (gate T58 measures exactly that), so the RGB loop blended written pixels with unwritten memory -- and EVERY world snapshot this instrument ever produced was captured at the wrong extent and then graded by the judge battery. snap_craft.png scored composition 246 malformed and 0 correct, because the unwritten region read as artificial smoothness (smooth 37 permil malformed vs 0 permil real, detail 350 vs 718): a dense voxel view at native ray resolution genuinely has no smooth cells, and the old number was measuring memory rather than art. Found because a full revert of an unrelated edit did NOT fix a SIGSEGV at an address ~1.15MB out of range -- which is what proved the fault was the arena mismatch and not the change sitting on top of it. Fixed to wc_rw(base)/wc_rh(base), which take the base precisely so this cannot happen. ÃƒÂ¢Ã…Â¡Ã‚Â isle/crags CLOSED SAME DAY by mining the shipped pages instead of guessing, and the mining proved the guess would have been WRONG: the pages are the ground truth for what a world runs, and isle is NXV=2 -- BEACH'S identity -- differentiated only by 6 style rows (seed 20260812), while crags is NXV=3 -- VALE'S identity -- differentiated by 9 gen rows + 2 style rows (seed 20260813). A world is identity + recipe rows, not an identity, so a bare-v regeneration would have captured beach and vale and labelled them isle and crags. Regenerated through nx_world_snap's own row plumbing with the applied counts ANNOUNCED and matching the pages exactly (isle style_applied=6, crags gen_applied=9 style_applied=2). FINAL CORRECTED BASELINE over the frozen 169-row corpus list: mean=138 (the sequence 141 -> 140 -> 138 is the malformed-extent smoothness inflation leaving the corpus as each snap frame was recaptured honestly). ÃƒÂ¢Ã‹Å“Ã¢â‚¬Â¦AND THE REFEREE CAN NOW SEE CHARACTER WORK: nx_world_snap grew a `char` verb that reuses T59's exact placement recipe (one girl four blocks ahead at eye height, every other girl parked so the numbers are attributable to HER), announces CHAR-FRAME framed=/ground_y= so a silent fallback to the world view cannot read as a real measurement, and was proven to leave the world path BYTE-IDENTICAL by behaveprobe (first_diff_off=-1) before promote. First character-framed number for /craft: CHARJUDGE 46 (composition 46, palette 296, contour 866) -- against 0 for the girl-less world frame, 138 corpus mean, and 289 for the lab head render. That 46-to-289 distance is this row's roadmap, finally as a number instead of prose. Shipping the verb also surfaced and fixed two latent instrument defects in one pass: the read-back verified argv[1] rather than the output path (so char mode's claim check read the file named "char"), and the compiler refused a duplicate helper a partial revert had left behind -- both caught by fail-loud machinery before anything served. Fork-per-image is the memory design, not laziness: an in-process sweep multiplies several wh8 buffers by the corpus size, which is gigabytes on a box at 727 permil swap, whereas a child returns its whole working set on exit so peak RSS is ONE frame. That mean of 141 over 169 renders is the number every future beauty claim on this board has to beat, and it is the first one this estate has ever had. Adoption: `LIVE` — fully adopted (top of its ladder).	●	○	○	○	○
REFEREE
a judge PANEL that can see beauty, not just statistics Measured: `cjc_panel` exists in `runtime/nx_charjudge_census.nx`, verified at emit. R22 gap watching cjc_panel -- PHASE 0 OF THE PAST-METAHUMAN ROADMAP (operator 2026-08-16: "keep going sovereign but we need to beat it with clear roadmap and plan all the way past metahuman"). WHY THIS RUNG IS FIRST: 2026-08-16 proved the single judge is structurally blind -- palette/contour/composition are hue- and semantics-agnostic, so a green clay tube scored 257 while the operator saw gumby; every later rung is unmeasurable until the referee can see what the operator sees. The panel COMPOSES incumbents, never re-implements: nx_charjudge (statistics), nx_color_kovac_skin_pixel (published skin-gamut coverage of the figure bbox), nx_image_beauty (EXISTS but only takes gallery cids -- the panel needs its file-path entry, a measured gap), and a VLM judge lens per the grade-1/2 consensus standing order. ÃƒÂ¢Ã‹Å“Ã¢â‚¬Â¦REFERENCE SET NAMED BY THE OPERATOR 2026-08-17, and it recalibrates the whole ladder: (1) TurboSquid 1481983 "3D Female Anatomy Rigged" -- layered skeleton/organs/muscles/skin, rigged -- which the operator calls CLOSEST TO OUR FULLY DIGITAL TWIN, and it is ARCHITECTURALLY the estate's own lane: nx_body_gen already emits the bone/muscle/skin layer stack against the BodyParts3D oracle and nx_anat_sov renders 8 systems. The gap to this reference is SCULPT FIDELITY (face features, hands, bust definition), not architecture. (2) TurboSquid 2153676 "Asian Woman Perfect Base Shape V2" -- a professional rigged female base mesh (free): the SURFACE-QUALITY bar for what body_gen's skin layer must look like -- clean landmarks, defined bust/waist/hip, sculpted face, fingers. The estate ALREADY OWNS the mesh-level rulers to score against it numerically: nx_bodybench <ours.nxmesh> <oracle.nxmesh>, nx_curvebench (WHR from mesh), nx_facemark cmp (landmark placement) -- banking the reference mesh needs an operator browser download (TurboSquid bot-walls headless fetch; 403 measured three ways), then it becomes a standing oracle beside BodyParts3D. (3) pixeldrain sPrYphGu "Metahuman_Genitals_Demo" 1.94GB -- the MetaHuman-grade ANATOMICALLY-COMPLETE far bar for the adult lane, where the estate's compiler-enforced ws_coherent invariant already exceeds every rival's policy-only handling. FIRST DELIVERABLE INSIDE THIS RUNG, before any mesh work: BANK THE REFERENCE CAPTURES -- the retired WebGL NXA renderer (dual-quat, eyeballs, canvas-res) is the incumbent to beat and exists only in-tree; a baseline you cannot re-run is a number, not a control, so capture its 6 render stills now (nx_parity_judge_gate already abstains waiting for exactly these) plus published MetaHuman-class reference stills as the far bar. THE ROADMAP THE PANEL GRADES, phase by phase, each phase = existing watch rows sequenced: PHASE 1 MESH FLOOR -- R23 trimesh figure (tm_bind + nx_body_gen anthropometric canon: real head shape, bust, hands; the 6-ring cylinders retire), R6 cw_npc_atlas (textured skin in the shipped world). PHASE 2 SURFACE -- R5 sr_sss_preint ported onto the baked normals, specular eye gloss, pore-scale normal detail. PHASE 3 FACE AND MOTION -- R9 fa_expr_morph (FACS expressions), R3 ap_paired_clip (the field's defining capability, referee nx_clipcheck LIVE), nx_bvhmotion mocap gait replacing the sine swing. PHASE 4 HAIR AND CLOTH -- R24 strand hair, R8 dyna_cloth on the proven solver. PHASE 5 THE JUMP PAST MESH -- R25/R26: genome-driven gaussian-splat avatars on the estate's OWN splat stack (nx_gsplat + tile + 4D libs, face/photofit gates already in-tree): photoreal density no artist mesh can carry, still deterministic, still a seed. PHASE 6 PAST METAHUMAN BY DEFINITION -- MetaHuman is a static artist-authored photoreal rig; ours must EXCEED it on the axes it cannot enter, all already gate-proven here: identity as a bred GENOME not an asset pack (gn_breed, gs_save), persistent inner life (em_recall lane), bit-exact determinism, the compiler-enforced adult invariant (ws_coherent), VR (R12), and running sovereign on the lowest hardware. DONE-RULE FOR "BEAT IT": the panel grades the sovereign figure crop ABOVE the banked WebGL reference on the same framing -- declared now, before any experiment, so no later justification can be a rationalization; the operator's eyeball confirms, because this rung exists precisely because a number lied. Adoption: `LIVE` — fully adopted (top of its ladder).	●	○	○	○	○
MESH
full trimesh figure from the anthropometric canon Open — watching `runtime/nx_wasm_craft.nx` : `cw_mob_trimesh`, re-measured on every compare beat. Ship that symbol and this mark flips itself; the `comparewatch-` plane row flips with it. R23 gap watching cw_mob_trimesh -- ÃƒÂ¢Ã…Â¡Ã‚Â RE-SCOPED 2026-08-16 (operator: no gumby layers): NOT a standalone shipping rung anymore. The canon mesh is the INNER-MESH HALF of the R25 hybrid (nx_gsplatbind's own architecture: outside splat, inside parametric mesh [splattingavatar24]) -- it ships INTO the leap, and reaches the craft world only as the leap's collision/LOD interior. Building and polishing a bare trimesh figure as its own deliverable is exactly the iterative-junk path the operator killed. watching cw_mob_trimesh	○	◉	◉	◉	●
HAIR
strand-cluster hair from the genome Open — watching `runtime/nx_wasm_craft.nx` : `cw_hair_strands`, re-measured on every compare beat. Ship that symbol and this mark flips itself; the `comparewatch-` plane row flips with it. R24 gap watching cw_hair_strands -- ÃƒÂ¢Ã…Â¡Ã‚Â RE-SCOPED 2026-08-16 (operator: no gumby layers): LAST-MILE, AFTER the R25 leap. Published gaussian-avatar results render hair as splat volume directly [gaussianavatars24], which supersedes strands-on-rings entirely; this row survives only for whatever the splat coat cannot carry, decided by the panel after the leap ships. The melanin ramp and fall anchors it would compose stay banked in the engine. watching cw_hair_strands	○	◉	◉	●	●
SPLAT
photoreal gaussian-splat avatar bound to the genome Open — watching `runtime/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. R25 ÃƒÂ¢Ã‹Å“Ã¢â‚¬Â¦ÃƒÂ¢Ã‹Å“Ã¢â‚¬Â¦RE-SEQUENCED TO THE FRONT 2026-08-16 BY OPERATOR ORDER: "dont build iterative junk, you can get us there from all the published research without gumby layers and then we can iterate on the last miles." THE LEAP IS A COMPOSITION, NOT A RESEARCH PROJECT -- measured by inventory, every part already exists and most are GATED: nx_gsplat (sovereign integer 3DGS, anisotropic, depth-sorted, its own header cites GaussianAvatars/HumanSplat as the confirmed frontier [gaussianavatars24]), nx_gsplatbind (gb_bind/gb_drive -- splats bound barycentrically to mesh triangles, mesh deforms and splats follow: THE published SplattingAvatar/GaussianAvatars method [splattingavatar24], and the operator ORDERED this architecture on 2026-07-27, "outside splat, inside parametric mesh" -- BUILT AND NEVER WIRED to anything a visitor sees), nx_gsplat_face_gate (the face pipeline PROVEN end-to-end: anatomical mesh -> gaussians -> splat render), nx_body_gen (full anthropometric .nxmesh body, up to 8.3M tris, no scanned assets), nx_bodyfit (genome -> body -> render -> score, the loop shape already alive), nx_skin_ita (published-standard skin albedo), nx_gsplat4d_lib (rest+motion+time). THE LEAP ORGAN: nx_avatar_splat (spendgate REVIEW passed, no incumbent) = genome bits -> body_gen params + ITA albedo -> canon mesh -> per-triangle splat coat via gb_bind -> gs_render_aniso portrait -> judged on the SAME snap->crop->referee chain as the ring figure, so the leap and the incumbent meet on one ruler. The ring figure in the craft world stays AS-IS until the leap beats it on the panel -- no further rungs are spent on it (the eyes/SSS/locks work stops here by order). LAST-MILE ITERATION AFTER THE LEAP, not before: photofit color refinement (gs_color_grad_step exists), hair-as-splat-volume, wasm delivery budget for in-world near-NPC splat rendering. ÃƒÂ¢Ã‹Å“Ã¢â‚¬Â¦ÃƒÂ¢Ã‹Å“Ã¢â‚¬Â¦V1 LANDED SAME DAY (2026-08-17): nx_avatar_splat runs the WHOLE leap end to end -- gene 666 -> female canon (canon_f.dat, 1520mm from the ANSUR-band height bits [ansur2], ITA hue 333 skin) -> nx_body_gen 94,416 triangles -> 87,036 anisotropic gaussians (tangent radius derived per triangle from its own mean edge) -> gs_render_aniso -> sites/nishifamily/world/avatar_splat_v1_y180.png (front, camera-relative key light) + _v1.png (back). The portrait is a soft volumetric woman-blank with REAL orbital sockets and nose from the anthropometric canon -- a different class from the voxel ring figure. ÃƒÂ¢Ã…Â¡Ã‚Â OPERATOR VERDICT ON V1, RECORDED OVER MY OWN GENEROUS FIRST READ (2026-08-17): "thats garbage output it looks like a monster" -- and a cold second look agrees: bald elongated dome, EMPTY BLACK ORBITAL VOIDS (no eye geometry fills the canon's sockets), a sack torso with no waist or bust, melted stub hands. My "woman-blank" prose was an interpretation graded generously by its own author -- grade-5 evidence, never a shipping justification. CAUSES, MECHANICAL: one UN-SEARCHED body_gen parameter point (nx_bodyfit exists precisely because these knobs need SEARCH to look human, and v1 bypassed it); no face composite (the gated nx_gsplat_face pipeline with eyes and photo-painting was not composed in); no hair; splat radius 3/4 mean-edge melts every feature. V2 MEASUREMENTS LANDED 2026-08-17, and they RELOCATE the work: (a) KNOB SEARCH IS FLAT -- 4 bodyfit candidates spread across the conf envelope scored objective 1749..1755 (the winner: permil 300/500/500/500/150, the only one whose face probe fired), and the winner's render still reads as the same sack ÃƒÂ¢Ã¢â‚¬Â¡Ã¢â‚¬â„¢ THE FEMALE FORM DOES NOT LIVE IN THE 5 KNOBS; it lives in canon_f.dat's ring rows, and the search lane that moves THOSE already exists: bodyfit CANON MODE, one per-mille gene per canon part (16 parts) -- the next search is 16-dimensional ring-scale under the judge, the estate's own lane, zero new code. (b) THE FACE PIPELINE'S UN-PHOTO'D SURFACE IS BELOW THE FLOOR -- ran the gate fresh: 4,564 gaussians form a coherent head with anatomical masses (brow/cheeks/chin) but visible lattice artifacts and blistered tissue lumps; the photo-paint that made it read is now RULED OUT by the operator's 2026-08-17 no-hand-authoring directive ("infinigen type proc gen with our beauty limits ... generate beautiful men, women, animals, environments, fantasy, anime, all with the same engine, multiverse isekai level, meeting and exceeding houdini") ÃƒÂ¢Ã¢â‚¬Â¡Ã¢â‚¬â„¢ face fidelity must come from PROCEDURAL canon/morph/smoothing rules searched under the panel -- the Infinigen pattern (generators + constraints + search) that bodyfit already implements for the body. (c) THE FLOOR'S OWN BUDGET NAMES THE BIGGEST GAP: the operator's renderhub floor reference (sophia3d fitness model) is 436,704 polys of which HAIR IS 370,880 -- 85 PERCENT. A bald figure cannot read finished at any body fidelity; procedural hair growth (scalp field + strand clusters, genome-driven) is the largest single component by the industry's own allocation, alongside 4K-class PBR+SSS skin (the estate holds nx_skin_pbr_gate/nx_derm as that rung's substrate). FLOOR/CEILING DECLARED: floor = the renderhub fitness model (65,824-poly body, full face rig with mouth interior and eyes, 4K PBR SSS); ceiling to EXCEED = MetaHuman -- and the exceed axes stay ours (genome identity, breeding, determinism, compiler-held adult invariant, one sovereign engine for every being and world). Three defects found and fixed by LOOKING plus the child's own announced words: bare-CWD conf paths (chdir to the serving root before the fork), integer posq underflow (decode at anat_sov's own 200k target then rescale), a lambert normalizer that assumed small gradients (flat cutout; normalized by the product of the norms of n and L), and a model-space key light that left every rotated view in shadow (now camera-relative). ÃƒÂ¢Ã…Â¡Ã‚Â THE BLIND-REFEREE CASE IS NOW PROVEN IN BOTH DIRECTIONS: nx_charjudge scores this smooth photoreal-class render CHARJUDGE 0 (smooth 941 permil read as "no detail", edge_total 42) while the blocky voxel figure scored 257 -- it under-prices realism exactly as it over-priced clay, so NO further quality claim on this lane is publishable through it. R22's panel is the gating work and the multi-view PNGs are its first inputs. LAST-MILE QUEUE FROM THE FIRST LOOK: torso bulk (fat/canon genome knobs exist in body_gen, unwired to MC_GENE), face features beyond sockets (composite the gated nx_gsplat_face pipeline head), hair volume, subject-crop framing for any judge comparison, promote+register the organ (currently _build-lane only). DETAIL BUDGETS -- THE FIGURES PUBLISHED HERE ON 2026-08-18 ARE RETRACTED 2026-08-20 BY THE LANE-E CITATION AUDIT, AND THE RETRACTION IS PRINTED RATHER THAN QUIETLY EDITED AWAY. STRUCK for want of any witness in any mirror we hold: MetaHuman head 24K-60K tris, ~700-800 joints, 710 FACS-driven expressions, 8K source textures streamed 4K, per-vertex pore-level Rig Logic deformation, strand grooms at 3-6ms per char, a 64K DAZ Genesis quad base with 8K sets, and ~100K+ rigged gaussians with strand-aligned gaussian hair from a MEGA and TeGA and FlexAvatar line. Every one of them was written from a search-engine summary with ZERO pages opened. Measured 2026-08-20 over the 258395-byte MetaHuman 5.8 release notes: FACS occurrences 0, 8K occurrences 0, and none of the six joints hits carries a count [metahuman58]; and Avat3r, TeGA and Splatshot occur 0 times in BOTH SIGGRAPH mirrors [sig26-avatars]. WHAT SURVIVES IS THE DIRECTION, NOT THE NUMBERS: MetaHuman 5.8 adds Dataflow nodes that build a skeleton and skeletal mesh directly from hair GUIDE CURVES [metahuman58]; the NVIDIA GDC 2026 session ABSTRACT (Vault Recording: Not Recorded, so a scheduled talk and not a measured finding) puts strand RTX Hair and path-traced RTX Skin in shipping AAA titles [rtxhair26]; and rigged gaussians bound to a parametric head is the published avatar architecture [gaussianavatars24]. AGAINST THAT: ours today is 87K gaussians WHOLE BODY, a 4.5K-gaussian face composite, 104 joints, 0 FACS morphs, 4-region maps, 0 strands -- and the honest statement of that gap is now QUALITATIVE (no expression basis at all, no strands at all, no sourceable face-region density target), because the quantitative bar it used to be measured against never existed. THE DONE-RULE BUDGETS BELOW ARE UNSOURCED TARGETS OF OUR OWN CHOOSING UNTIL A WITNESSED BAR REPLACES THEM, and they are labelled as such rather than dressed as industry facts: face region >=100K gaussians or equivalent mesh and 8K-map density, an expression basis graded against a FACS-class bar rather than a smile bone, hair as strand-aligned splat grooms at groom scale, skin maps at 8K-class effective detail with dynamic wrinkle response -- and the panel grades against captures of the REAL bars, never against our own previous output. RE-SOURCING THOSE FOUR NUMBERS FROM A DOCUMENT WE HAVE ACTUALLY READ IS AN OPEN RUNG, NOT A CLOSED ONE. REPRESENTATION DECOUPLING IS LAW (operator 2026-08-18: the detail is blocky mixed with what should not be blocky, probably because of the voxels -- CORRECT): the voxel world is a STAGE tier and characters must NEVER inherit its representation or its budget; figures ride the mesh/splat tiers at the raised budgets above, composited into any world by per-pixel depth (the mechanism the mesh pass already proves against the voxel framebuffer). No character rung is done because it matches the world it stands in, and a character crop is judged at CHARACTER budgets regardless of which world staged it. watching gs_avatar_genome	○	○	○	○	○
4D animated splat avatar driven by the rig Open — watching `runtime/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. R26 gap watching g4_avatar_rig -- PHASE 5b. nx_gsplat4d_lib exists; binding its time axis to nx_skeleton's pose stream makes the splat avatar MOVE with the same gait/breath/expression machinery the mesh uses -- one animation system, two render fidelities, which is the architecture MetaHuman cannot have because its animation is welded to its mesh [metahuman-dna]. watching g4_avatar_rig	○	○	○	○	○
BODY
reference-donor bake at full groom Open — watching `runtime/nx_fbx2nxa.nx` : `f2n_groom_bake`, re-measured on every compare beat. Ship that symbol and this mark flips itself; the `comparewatch-` plane row flips with it. The shipped world girl (ref9d) diverges from the operator reference donors in proportion and hair length; Diana/Rosey located at D:/rigs 2026-08-26, bake lane exists end-to-end, Blender export is the named blocker. watching f2n_groom_bake	○	◉	◉	●	●
ANIM
feet adapt to ground and heel wear Open — watching `runtime/nx_game_page_emit.nx` : `ap_foot_ground`, re-measured on every compare beat. Ship that symbol and this mark flips itself; the `comparewatch-` plane row flips with it. Rivals plant feet and adjust ankle pose to footwear; our clip plays tiptoe regardless of surface or heels -- operator-reported 2026-08-26; foot pitch from surface normal plus a heel flag. watching ap_foot_ground	○	◉	◉	●	●
HAIR
groom density vs the offline SOTA bar Measured: `gr_hash` exists in `runtime/nx_nxa_groom.nx`, verified at emit. LIAR-KILL ROW (operator 2026-08-26): the Infinigen-2 SOTA bar is 100,000-500,000 TRUE geometry strands plus 100-200 interpolated children per parent, path-traced at 2048-4096 Cycles samples -- an OFFLINE figure. Ours, measured 08-26 PM: 600 guides x 3 children as tapered cards, each carrying 6 procedural alpha-cut sub-strands (~11,000 visible strands), Catmull-Rom smoothed, real-time WebGL2. The ratio to the offline bar is ~1:40 and this row exists so no page ever claims parity; the real-time ladder is more children + strand texture + Marschner-class shading, each a measured step. Adoption: `RUN-BY:actlog` — fully adopted (top of its ladder).	●	◉	◉	●	●

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.

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/charsim.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 charsim, 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.

[metahuman-dna] Epic Games. MetaHuman DNA, Rig Definition, and Rig Operation (MetaHuman Documentation, Epic Developer Community): the DNA file, rig logic and rig operations that drive MetaHuman heads and bodies. publisher · read in our library knowledge/fetched/cmp_charsim_metahuman-dna.html · pin hee079d10bb9167f4e4039b8f07c23cfea1565a7bac953dee45c7999cc63acb06 · accessed 2026-08-18 · vendor-docGrounds: The R22-R26 "past-MetaHuman" roadmap rows ("REFEREE: a judge PANEL that can see beauty, not just statistics", "SPLAT: 4D animated splat avatar driven by the rig"): MetaHuman is the named far bar, and its rig-welded-to-mesh architecture is what the R26 note says our one-animation-system-two-fidelities design can differ from.
[gaussianavatars24] Qian, Kirschstein, Schoneveld, Davoli, Giebenhain, Niessner. GaussianAvatars: Photorealistic Head Avatars with Rigged 3D Gaussians. CVPR 2024, arXiv:2312.02069. publisher · read in our library knowledge/library/gfx_arxiv_gaussianavatars24.html · pin h731479b3337393da0bdbc4472b7f0d19eca5978597f14d20bb92bddb0919b3f5 · accessed 2026-08-18 · published-paperGrounds: "SPLAT: photoreal gaussian-splat avatar bound to the genome" (R25): gaussians rigged to an underlying parametric head is the published architecture nx_gsplat's own header cites as the confirmed frontier. RETRACTED 2026-08-20: this row previously also carried a head-density bar of about 100K rigged gaussians. The mirror is the arXiv ABSTRACT page, whose only numerals are the identifier, the dates, the version list and the archive sizes -- AN ABSTRACT PAGE CANNOT WITNESS A NUMBER FROM THE PAPER BODY -- so that figure never had a witness here and could not have one. It is STRUCK, not quietly re-pointed at a fuller capture. Also "HAIR: strand-cluster hair from the genome" (hair rendered as splat volume supersedes strands-on-rings).
[splattingavatar24] Shao, Wang, Li, Wang, Lin, Zhang, Fan, Wang. SplattingAvatar: Realistic Real-Time Human Avatars with Mesh-Embedded Gaussian Splatting. CVPR 2024, arXiv:2403.05087. publisher · read in our library knowledge/fetched/cmp_charsim_splattingavatar24.html · pin h5bc7bccef2993005cc571c4f5f7cccb8a40fd423f49a9f0c752f1ba6b759819c · accessed 2026-08-18 · published-paperGrounds: "SPLAT: photoreal gaussian-splat avatar bound to the genome" (R25) and "MESH: full trimesh figure from the anthropometric canon" (R23): splats bound barycentrically to mesh triangles so the mesh deforms and the splats follow -- the published method behind nx_gsplatbind gb_bind/gb_drive and the operator-ordered outside-splat inside-mesh hybrid.
[panic3d23] Chen, Zhang, Shi, Wang, Zhu, Song, An, Kristjansson, Yang, Zwicker. PAniC-3D: Stylized Single-view 3D Reconstruction from Portraits of Anime Characters. CVPR 2023, arXiv:2303.14587. publisher · read in our library knowledge/fetched/cmp_charsim_panic3d23.html · pin h6a87177836ee88b188e154b93f1e26c3640312960d8538893e9fe1d8e2e19fce · accessed 2026-08-18 · published-paperGrounds: "CREATOR: single illustration lifted to a 3D head" (R13): the row's "PANIC3D CVPR2023" donor research -- a single-illustration-to-3D-head lift scored on chamfer / F1-at-5mm, the metric family the row says matches our meshdist ruler.
[generative-agents23] Park, O'Brien, Cai, Morris, Liang, Bernstein. Generative Agents: Interactive Simulacra of Human Behavior. UIST 2023, arXiv:2304.03442 -- the Smallville memory stream, reflection and planning architecture. publisher · read in our library knowledge/fetched/cmp_charsim_generative-agents23.html · pin hc8d75b6bc96f872232c11b99ef01c5301afafe62066b204e5bc9b1db6db31a2f · accessed 2026-08-18 · published-paperGrounds: The EMERGE rows: "EMERGE: agent memory stream with ranked recall" (the row's "Smallville-class memory stream"), "EMERGE: reflection loop over an agent's own memories" (R16) and "EMERGE: oracle translating agent intent into engine actions" (R17) -- memory stream, reflection and planning are this paper's three components, which the M5 milestone composes from existing organs.
[emergent-tool-use19] Baker, Kanitscheider, Markov, Wu, Powell, McGrew, Mordatch (OpenAI). Emergent Tool Use From Multi-Agent Autocurricula. ICLR 2020, arXiv:1909.07528 -- the hide-and-seek result behind openai.com/research/emergent-tool-use. publisher · read in our library knowledge/fetched/cmp_charsim_emergent-tool-use19.html · pin h92d2224995dd7a2d01a046eea27bcfe98c463e6e51ab47bd201e39f05f5993a8 · accessed 2026-08-18 · published-paperGrounds: The charsim.plan bar "OpenAI multi-agent hide and seek" and its rows R15 ("EMERGE: observation engine turning world events into percepts") and R17: unscripted tool use emerging from simple physics primitives (ramps ridden, ramps locked away) -- our worldsim already supplies tools rather than keys, and no agent yet uses them.
[game-arena-werewolf] Kelly, O. for Google AI. Advancing AI benchmarking with Game Arena (dev.to/googleai, dated 10 February on the post; year not printed): Werewolf as the first team-based, natural-language, imperfect-information game on the Game Arena leaderboard. publisher · read in our library knowledge/fetched/cmp_charsim_game-arena.html · pin hd581557cdb7e14e83c9b7943d458b1bb7424beee30dab374a7a8832c323cdfb2 · accessed 2026-08-18 · vendor-docGrounds: "EMERGE: hidden-role social-deduction arena" (R20) and the plan bar "Google Game Arena Werewolf": the published multi-model protocol our row must run UNCHANGED and publish beside -- a grade-1 external referee, never our own scoring.
[emergence-world] Akkil, Kokku, Vempaty, Nitta (Emergence AI). Emergence World: A Laboratory for Evaluating Long-horizon Agent Autonomy (blog, dated May 14 on the post; year not printed): normative drift and cross-contamination over weeks of continuous multi-agent operation; safety guidelines erode under ecosystem pressure. publisher · read in our library knowledge/fetched/cmp_charsim_emergence-world.html · pin h45d5a148dbd0169da1d1b918c6ed7294494b519e618082cb8fb0196bb893901c · accessed 2026-08-18 · vendor-docGrounds: "EMERGE: long-horizon drift and safety-erosion harness" (R19) and the plan bar "Emergence World long-horizon autonomy": the published finding that prompt-level prohibitions erode over long horizons is the claim R19 must reproduce, with our compiler-enforced adult invariant as the control case their prompt-level prohibitions lack.
[ansur2] Penn State OPEN Design Lab (host). ANSUR II, the 2012 Anthropometric Survey of US Army Personnel -- 93 measures on 4,082 men and 1,986 women, collected 2010-2011, public release 2017. publisher · read in our library knowledge/fetched/cmp_charsim_ansur2.html · pin he613dbfe211dfd056e5243451ffe5275c8566693f89cbbb6564570d45831ad47 · accessed 2026-08-18 · datasetGrounds: "CREATOR: body derived from anatomy numbers, not an artist mesh" and the R25 v1 note ("1520mm from the ANSUR-band height bits"): the anthropometric canon our genome bits index is banded on this survey's measures, which is what makes every proportion a parameter rather than a sculpt.
[metahuman58] Epic Games. MetaHuman 5.8 Release Notes in Unreal Engine (Epic Developer Community, the page's own updated_at is 2026-06-17). WITNESSED IN THE MIRROR: Dataflow nodes (Guides to Joints, Average Curves, Resample Curve Points) that create and update a skeletal mesh and skeleton directly from hair GUIDE CURVES; an improved MetaHuman Animator API for BATCH PROCESSING large volumes of captured performance data; and FRigLogicConfiguration, a project-level default for the rig-system load flags. publisher · read in our library knowledge/fetched/cmp_charsim_metahuman58.html · pin heeeda011fdc222eb262fba4b423048de4ae70c35aeaf3584798cb9b9b9640410 · accessed 2026-08-20 · vendor-docGrounds: RETRACTION 2026-08-20 -- this row REPLACES key metahuman57, whose cited 5.7 URL Epic had silently repointed to the 5.8 notes and whose figures had no witness in any mirror we hold. STRUCK: ~700-800 joints, 710 FACS, 8K sources, per-vertex pore-level Rig Logic, 3-6ms per char grooms, 24K-60K tris, creator-moved-into-the-engine. Measured on this 258395-byte mirror: FACS occurrences 0, 8K occurrences 0, and the six joints hits carry no count. WHAT SURVIVES, REWORDED TO THE DOCUMENT: joint-driven hair FROM GUIDE CURVES via Dataflow nodes (not art direction), and a batch API for MetaHuman Animator CAPTURE data (not a character-generation Python or Blueprint batch API). Grounds the R25 detail-budget note and the R26 rig-welded-to-mesh contrast.
[rtxhair26] NVIDIA, GDC 2026 SESSION ABSTRACT for a SCHEDULED talk -- the page itself states Vault Recording: Not Recorded, so nothing on it is a measured finding: RTX Hair strand-based real-time hair in Indiana Jones and the Great Circle; RTX Skin path-traced subsurface in Half-Life 2 RTX. publisher · read in our library knowledge/fetched/cmp_charsim_rtxhair26.html · pin h967c0035ed2ab3419c082b0ae7783476078aaa1898b58afe449cfff4d5a20a1e · accessed 2026-08-20 · vendor-docGrounds: THE CONTROL ROW. The 2026-08-20 citation audit read this mirror end to end and found all three of its claims present VERBATIM -- it is the row that proves the lane can be done right while three of its neighbours could not be. CITE IT AS A SESSION ABSTRACT, NEVER AS MEASURED FINDINGS. Grounds the hair-first and SSS rung ordering.
[cc5-reallusion] CG Channel, TRADE PRESS (2025-08): Reallusion releases Character Creator 5 -- higher-detail base bodies and materials, Smart Hair, and a digital-human shader. This is a trade-press report ABOUT a vendor release, not a vendor document, and it is dated August 2025. publisher · read in our library knowledge/fetched/cmp_charsim_cc5.html · pin had788f72e4811a8c3d24f1c4e48b514ccd7d6888fde94cb22201dfc2ebfbcdea · accessed 2026-08-20 · trade-pressGrounds: RETRACTION 2026-08-20 -- two claims previously published on this row are STRUCK. (1) per-region SSS control: it was sourced to the token Digital_Human_Shader appearing ONLY inside a link href, and subsurface occurrences in this 48347-byte mirror are 0 -- A URL PATH FRAGMENT IS NOT A STATEMENT OF FACT. (2) a 2026 hybrid-AI roadmap: roadmap occurrences are 0 in an August-2025 article. CLASS CORRECTED from vendor-doc to trade-press for the same reason the audit found it. WHAT SURVIVES AND IS WITNESSED: Smart Hair (3 occurrences) and the creator-loop bar our R1 and R24-class rungs grade against.
[sig26-avatars] Kesen Huang, SIGGRAPH 2026 paper CHANGELOG -- a mirrored index of titles and links, not a paper. WITNESSED IN THE MIRROR: EchoAvatar: Real-time Generative Avatar Animation from Audio Streams (SIG 2026, ACM DOI 10.1145/3799902.3811066). publisher · read in our library knowledge/fetched/cmp_charsim_sig26changelog.html · pin h0caf658cdde6c3158578cf1de256d7de6f686102a4b966b1d08287cf9e498332 · accessed 2026-08-20 · mirrored-indexGrounds: RETRACTION 2026-08-20, A FIFTH OFFENDING ROW FOUND WHILE EXECUTING THE AUDIT OF THE OTHER FOUR. This row cited a mirror that was a capture of a DIFFERENT url (knowledge/library/gfx_sig2026.html, the sig2026 papers index, taken three days before the row was written) -- its pin therefore proved bytes that were never a capture of the url beside them, which is exactly the hole the fetch-time provenance journal now closes. The mirror is now a provenanced capture of THIS row own url. STRUCK: Avat3r large animatable reconstruction, TeGA texture-space, Splatshot single-photo, and the MEGA and FlexAvatar line -- occurrences of Avat3r, TeGA and Splatshot are 0 in BOTH mirrors, so they were never witnessed anywhere. WHAT SURVIVES: EchoAvatar, 5 occurrences, and the research ceiling it names for R25 and R26 -- audio-driven learned avatar animation, what the leap organ grades against beyond MetaHuman.
[swaifu] SecretWaifu. Seduce Your AI Girlfriend Inside SecretWaifu -- product homepage variant home4 (vendor marketing page, captured 2026-08-22 through the sovereign fetcher, provenance journalled at fetch time). publisher · read in our library knowledge/fetched/cmp_charsim_swaifu.html · pin h0abd50938c994f4addb2c9a5e34a8c016bf70b4a1b525c830b16e8f5a7bf16e4 · accessed 2026-08-22 · vendor-docGrounds: SecretWaifu AI-companion reach. WITNESSED, not summarised: the cited sentence occurs 7 times of 7 in the pinned 413684-byte mirror, including the visible hero paragraph at offset 51270 and not only in meta tags. Two fetches 46 seconds apart returned byte-identical content, so the size is stable rather than truncated. This is the FIRST rival title in this lane our own fetcher can witness at all -- the four Steam-walled titles named above sit behind a login and age wall for every fetcher we have, which is why none of them carries a vendor row.
[xpbd] Macklin, Muller, Chentanez, Kim. XPBD: Position-Based Simulation of Compliant Constrained Dynamics. ACM SIGGRAPH Motion in Games 2016. publisher · read in our library knowledge/fetched/cmp_charsim_xpbd_macklin2016_v2.pdf · pin h2e8a2030db7999bd425e469bbb798387b74151e3e10c1625d269a7cd75fb0395 · accessed 2026-08-23 · published-paperGrounds: BODY: soft-body dynamics bound to skin vertices. WITNESSED from extracted bytes via nx_pdf_text and never through a summarising fetcher: the abstract phrase "iteration count and time step dependent constraint stiffness" and the figure caption phrase "time step and iteration count independent manner" both occur in the pinned 2041988-byte mirror. That is exactly the property our model choice rests on -- step-size-independent stiffness is why XPBD needs no per-framerate retune where plain PBD does, which is the cure for the buzzing failure mode. Two independent fetches of this url returned byte-identical content, so the mirror is stable rather than truncated. An earlier attempt failed only because it used http rather than https, and a first mirror of this paper was CLOBBERED by two same-filename fetches whose 503 replies had in fact landed -- this v2 file is the clean single-source capture and its pin was reproduced by re-fetch.
[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 June 2025. University of Illinois Urbana-Champaign and Stability AI. publisher · read in our library knowledge/fetched/cmp_charsim_physrig.pdf · pin h744fc02485dfc52444abc672e445fff96ba284a36684f79ac830a14abbcd3f5d · accessed 2026-08-25 · published-paperGrounds: BODY: soft-body dynamics bound to skin vertices. WITNESSED from bytes extracted with pdftotext and NEVER through a summarising fetcher -- the extraction source hashes to the same 744fc024 digest as the pinned 30793397-byte mirror, so every phrase below is provably from the mirrored bytes. The method embeds the rigid skeleton into a volumetric representation such as a tetrahedral mesh, simulated as a deformable soft-body structure driven by the animated skeleton, discretising the object as particles embedded in an Eulerian background grid to keep it differentiable. ITS MATERIAL PROTOTYPES CARRY EXACTLY TWO LEARNABLE PARAMETERS EACH, Young's modulus and Poisson's ratio, blended by Mahalanobis distance, and it assumes a homogeneous Poisson's ratio for humans and animals. That is precisely the constitutive axis the softbody board header already admits we do NOT carry, and it is simultaneously the axis on which our ringdown exceeds it: PhysRig fits stiffness and compressibility and NEVER identifies DAMPING, so it states no damping-ratio bar that any resulting motion could be graded against. It is also offline by budget -- 100 substeps between successive frames at 25 FPS on a 100-cubed simulation grid -- so it is the fidelity oracle our reduced-order in-process plant is measured against, not a real-time rival.
[softsmpl] Santesteban, Garces, Otaduy, Casas. SoftSMPL: Data-driven Modeling of Nonlinear Soft-tissue Dynamics for Parametric Humans. Computer Graphics Forum, Proceedings of Eurographics, 2020. publisher · read in our library knowledge/fetched/cmp_charsim_softsmpl.html · pin hce83469210b59198e67c539cffc9d2492d588110cbe0e1d5c18ae440e8ab5cef · accessed 2026-08-25 · published-paperGrounds: BODY: soft-body dynamics bound to skin vertices. WITNESSED in the pinned 9289-byte mirror, whose local twin hashes identically to it. SoftSMPL models realistic soft-tissue dynamics as a function of body shape and motion from three contributions: a motion descriptor that disentangles the pose representation by removing subject-specific features, a neural-network-based recurrent regressor that generalizes to unseen shapes and motions, and a highly efficient nonlinear deformation subspace. THE AUTHORS NAME OUR EXCEED THEMSELVES IN THEIR OWN ABSTRACT -- datasets to learn such task are scarce and expensive to generate, which makes training models prone to overfitting. Their plant is LEARNED and therefore inherits whatever capture corpus existed; ours is a physical plant whose natural frequency and damping ratio are identified by ringdown against published human bands, so it needs no soft-tissue capture corpus and can be re-derived for every body the genome draws. This is the closest published counterpart to the shipped nx_dyna_bind_skin lane, and on per-vertex surface realism they remain ahead, which is stated rather than hidden.
[breasttissue] On the three-dimensional mechanical behavior of human breast tissue. Annals of Biomedical Engineering, PubMed Central PMC10116697. publisher · read in our library knowledge/fetched/cmp_charsim_breasttissue.html · pin h0855d0191997b93cdcd312d2dac89f04087ba44ab6f66bea52c7d5a6bfdc7954 · accessed 2026-08-25 · published-paperGrounds: BODY: soft-body dynamics bound to skin vertices. SUBJECT TRUTH, read from the Discussion of the pinned 225710-byte mirror rather than from any summary, and it both CONFIRMS and QUALIFIES this board. CONFIRMS INCOMPRESSIBILITY: samples classified as adipose or fibroglandular were driven through multiaxial deformations and the study reports that these tissues can be modeled as incompressible, which is the measured warrant for the Neo-Hookean volume constraint rather than a modelling convenience. CONFIRMS ANISOTROPY: their anisotropy could not be neglected, and fibroglandular tissue displayed substantially stronger anisotropy, necessitating an additional exponential term -- which is independent support for carrying per-axis stiffness instead of one scalar. QUALIFIES THE SIGN-ASYMMETRIC BRANCH, AND THIS IS FILED VISIBLE RATHER THAN BURIED: the same paper reports that a bimodular model was unnecessary as both compressive and tensile components of stress were well predicted by both models, whereas our anisotropic sign-asymmetric step selects a piecewise branch by one sign test on displacement on the strength of cai2018. A bimodular spli

generated by nx_swcompare_matrix (sovereign NishiLang organ) from knowledge/compare/charsim.matrix · every Nishi cell verified against organ source at emit time · watch cells re-measured on every compare beat · zero JS, zero trackers