VRM Seed-san - the donor format, into and out of the nishi ecosystem

/exceed showcase · format=vrm · slug=seed-san · every number re-derives from its evidence path; tiers per the three-tier display doctrine.

INBOUND — their original, into the nishi ecosystem

T1 ORIGINAL · PARTIAL

Seed-san by VirtualCast, Inc. (credit required and given; VRM Public License 1.0) - full VRM 1.0 avatar served as-is, 10.9MB; drag to rotate

their original, served byte-identical; our interop viewer renders mesh+skin - MToon materials, springbones and expressions are named T1 rungs, not yet rendered

T2 MIGRATED · PARTIAL-EARNED

OUR being wearing HER measured proportions AND her measured bent-elbow A-pose, fitted by the closed loop (nx_fitloop, 24 audited trials, -23pct) - drag to rotate, full interactivity restored. The INTERACTIVE deviation heatmap (drag-rotate the painted gap itself): /exceed/vrm/nishi_seedsan_heatmap.glb ; stills: nishi_fitted_seedsan.png + nishi_seedsan_heatmap.png

their DATA drives OUR rig, now through the CLOSED LOOP: 5 skeleton axes converged (worst 1.0pct, from the served glb); pose fitted BY MEASUREMENT (her bind pose traced as bent-elbow A-pose from her own mesh; nx_fitloop 24 audited trials, whole-body objective -23pct, completeness 84.4->75.8mm); L0 verdict HUMANOID; per-part table live (spine 5.2mm ... fingertips ~300mm); deviation heatmap published beside the render. OPEN and measured: head+hair geometry (magenta), arm length +93mm, per-part correspondence, UV/texture

T3 IMPROVED · OPEN
T3 bars NAMED BY US, measured never asserted (piecewise-parity program 1785973794): VISUALS per-part detail-variance parity + palette >=217 + UV/texture; PERFORMANCE quality-per-byte (their body 10.9MB/34k tris = the ratio to beat); CAPABILITIES superset = walk-in-file (ours, they lack) + interactive deviation heatmap (ours alone today) + 18 expressions (their bar) + reshape-with-reexport (the empty cell no surveyed tool occupies). Improved is a NUMBER on each bar

OUTBOUND — ours, into their world

T1 ORIGINAL · LIVE

our original: the nishi being in our own format (ORIGINAL license, every bit nishi)

our native artifact in the same viewer - one door, two directions

T2 MIGRATED · OPEN

OUR being wearing THEIR measured proportions - the R3 retarget applied (drag to rotate)

SLOT CORRECTION (appended, last-wins): the retarget is INBOUND (their data into our rig) and now renders there. OUTBOUND stays open: nishi -> VRM emit is the give-back path (C3 rung + nx_vrm_reshape 1785877444); Seed-san is its round-trip oracle

T3 IMPROVED · OPEN
improved-nishi bars as above; anatomy/motion rungs ride the mh lane

Capability radar — ours (filled) vs best-of-breed (outline)

UV-emit 0 vs 1 VRM-emit 0 vs 1 rig-joints 104 vs 55 reshape 1 vs 0 accessors 16 vs 389 expressions 0 vs 18

every axis scaled to its own stated max (manifest data, never a code constant); values clamp at the rim

Judge measurements

organaxisvalueunitevidence
nx_modelfactsglb_accessors389accessorslive sweep 2026-08-04 knowledge/status/exceed_vrm.list
nx_modelfactsglb_skinned1fileslive sweep 2026-08-04 (full humanoid rig)
nx_modelfactsglb_vrm1fileslive sweep 2026-08-04 - VRM detection positive on a VRM 1.0 file
nx_modelfactsglb_uv1fileslive sweep 2026-08-04 (UV-textured - the axis our exporter measures 0/12)
retarget-CORRECTEDshoulderW_ratio0.355theirs/oursR3 mapping v2 2026-08-04: SUPERSEDES the earlier 0.329 row - our root sits at y=0 but the hip joint at -40, so above-hips shoulder height is 560 not 520; the first denominator inflated every ours figure 7.7pct
retarget-CORRECTEDhipW_ratio1.202theirs/oursR3 mapping v2 - supersedes 1.116
retarget-CORRECTEDthighLen_ratio1.042theirs/oursR3 mapping v2 - supersedes 0.968
retarget-CORRECTEDupperArmLen_ratio1.370theirs/oursR3 mapping v2 - supersedes 1.273; fx1024 vector = shw363 hipw1231 thigh1067 arm1403
retarget-CONVERGEDcalfLen_ratio1.0002theirs/oursclosing measurement 2026-08-05 FROM THE SERVED GLB: their foot.L/shin.L = 0.38811/0.34247 = 1.13327 vs our Lank/Lkn = 460/406 = 1.13300; calf applied fx1024=1179, derived frame-cancelling via the CONVERGED thigh axis (ankle landmark joints added - calf now expressible)
retarget-CONVERGEDmax_axis_residual1.0pctratiore-measured 2026-08-05 from the SERVED glb, now FIVE axes (shoulderW/hipW/thighLen/upperArmLen/calfLen); calf residual 0.02pct - the closing measurement of debt 1785900993, every measured axis expressed in the 11-joint export
nx_retarget_showcasenative_render_verts27134vertsall-native emit 2026-08-05 (54248 tris, ovf=0 - refuses to render a truncated body; calf 1179 applied, the longer lower leg adds surface)
nx_bodybenchform_bodygen_baseline232permil2026-08-05 pm vs the CLEAN body oracle (34382 tris: robo_arm mesh + backpack/plug/cable prims 11-16 excluded by the new per-prim report+skip lane): body_gen default = front 598 / side 478 / shape 912 / detail 232. Her CLEAN-fitted profile (mean_abs_dev 151, was 211 contaminated) drives side to 585 but front 299: the radial fit ingests A-POSE ARMS into torso stations - per-part fit (nx_profile_fit part_id arg) is the named next rung
nx_bodybenchform_headline278permilFULL MATRIX vs the CLEAN body oracle 2026-08-05 late: atlas T2 being (retargeted) = 278 (front 616/side 442/detail 278) BEATS body_gen male-tuned 232, male-canon+her-profile 199 (front 611/shape 923 best-of-tubes), female-canon+profile 175. The retargeted proportions ARE her gross form. ⚠RULER GAP EXPOSED: the winner has NO ARMS - silhouette IoU + normal-variance cannot see missing limbs (the historical silhouette-rewards-width lesson recurring) - a structural limb-presence judge is the instrument's next tooth before any T2-swap decision
nx_meshdistform_gap_symmetric_mean_um74800umFIRST METRIC MEASUREMENT 2026-08-05 (NEW nx_meshdist: bidirectional point-to-surface, METRO/DTU accuracy+completeness, 10um quanta, self-floor 20um on 100mm tris / ~2um body-scale, overflow tooth bitten+fixed): atlas pawn = acc 41.5mm / COMPLETENESS 108.2mm HD95 412mm (her arm points have NO pawn surface within 40cm - the metric ruler SEES the missing arms that IoU could not); male-canon+profile tube = acc 63.1mm / compl 84.4mm HD95 334mm (arms present, POSE wrong: hanging vs her T-pose). Opposite failure modes now separable; next collapse = pose-matched candidate (T-pose the generate)
nx_bodypartsL0_taxon_verdictPAWN=NOT-HUMANOID(missing arms) / TUBE=HUMANOID / HER=HUMANOIDverdictSEMANTIC LADDER L0 LIVE 2026-08-05 (new nx_bodyparts, band-cluster structure census, gate 3/3 non-vacuous): her body = arms 2/legs 2/neck+head, shoulder 117 neck 14 bins; the retargeted atlas being = arm_clusters_max 1 = ARMS ABSENT, formally refused as a humanoid. The instrument now says what the eye said and NAMES the missing part - a metric score can no longer rank an armless body first
nx_kindL0_kind_verdictHER=humanoid(3/3) / OURS=NO-KIND-MATCH(first_fail arms-clear-of-torso)verdictOMNI ONTOLOGY LIVE 2026-08-05: kinds are PLANE ROWS not code (nx_kind, gate 4/4 incl a quadruped whose stack axis is DECLARED IN DATA). Her body matches humanoid on all 3 predicates; the retargeted being matches NO registered kind and names why. ⚠her mesh also satisfies the loose quadruped rules = the ruler needs specificity (most-rules-wins today; discriminative scoring is the named rung)
nx_gltf2meshL2_named_parts74partsSEMANTIC LADDER L2 LIVE 2026-08-05: the donor SEGMENTS ITSELF - JOINTS_0 resolved through the owning skin to NODE indices gives 74 named parts from 132 joints (hips, hand.L, breast.L/R, foot_rope_L.001 ...), emitted as NXMSH2 layers with the ORIGINAL author's names. No clustering heuristic, no invented vocabulary. Geometry unchanged: 34382 tris, still verdict=humanoid 3/3
nx_meshdistper_part_tablespine 5.2mm / chest 9.1mm / shoulder 10mm / hips 12mm / head 15mm vs fingertips 357mm / toes 305mmmmPER-PART COMPLETENESS LIVE 2026-08-05 (meshdist parts verb over the 74 donor-named layers): the TORSO CORE IS ALREADY 5-17mm - and the error grows MONOTONICALLY along each kinematic chain (shoulder 10 -> upper_arm 20 -> forearm 51-66 -> twist 154-184 -> hand 241-259 -> fingertips 344-357; thigh 19-24 -> shin 64-68 -> foot 278-305). A gradient that grows with distance from a joint is the SIGNATURE OF POSE MISMATCH ABOUT THAT JOINT, not form error - the pose-match rung is now MEASURED as binding, no longer hypothesis. Hair tail 7->189mm monotonic = geometry we simply do not have
nx_meshposepose_iteration_tablehang/T90/A40/A31/elbow40: hand 241/358/258/191/181mm, twist 154/282/237/164/122mmmmPOSE LOOP MEASURED 2026-08-05 (new nx_meshpose report->apply->parts, gate 4/4, rotation verified analytic): T-pose assumption FALSIFIED - her bind pose is a BENT-ELBOW A-pose (upper arm near-vertical, forearm out ~40-50deg, measured by tracing her own arm clusters). NO single-pivot rotation dominates all parts = a one-joint model cannot fit a two-joint chain; elbow-only wins mid-chain, whole-body completeness 84.4->75.8mm HD95 334->296mm. Residual = multi-parameter search (shoulder/elbow angles + elbow height + our arm 93mm LONGER than hers 725v632) = the nx_fitloop keystone 1785967773 with the per-part table as its objective
nx_fitlooppose2_convergeds=16deg e=90deg, objective 15.22M->11.76M um-sum (-23pct) in 24 audited trialsumTHE CLOSED LOOP RAN 2026-08-05: nx_fitloop (generator nx_meshpose <- judge nx_meshdist parts, coordinate descent, vacuous-objective guard) optimised the candidate pose UNATTENDED with every trial logged. Trace lessons: the s=0->16 step bought 3.4M (94pct of the gain); e beyond 16 rode a near-flat gradient to its bound because the forearm partially exits the elbow window at s=16 - A NEAR-INERT PARAMETER WITH A MONOTONE CRUMB TRAIL DRAGS THE SEARCH TO ITS BOUND (region windows must track the chain, the named v2 rung). Metrology-in-the-authoring-loop is now a working organ, not a program
nx_meshdistdeviation_heatmapgreen 211 / yellow 110 / orange 97 / red 77 / magenta 502 permil (10mm bands)permilTHE SOVEREIGN DEVIATION HEATMAP LIVE 2026-08-05 (nx_meshdist paint verb, ZEISS-Inspect-class): fitted candidate painted by distance-to-her-body, rendered by the existing viewer (per-tri colour was already in NXMSH2). THE PICTURE = the per-part tables at a glance: legs GREEN <10mm, chest core close, head+shoulders+arms MAGENTA >40mm. Histogram printed WITH the paint so picture and numbers can never disagree. Tooth: R/B channel swap in the render path (yellow/red band as cyan/blue) - ordinal readability intact, fix filed
nx_meshviewchannel_swap_fixedR-LOW pack restoredfixEVERY nx_meshview render had R/B swapped (packed R-high vs write_png r+g*256+b*65536 R-low) - flesh rendered blue for the lane's whole life and NOBODY COULD SEE IT until the heatmap, the first render whose colours MEANT something. Fixed+promoted 2026-08-05; heatmap now true green->yellow->orange->red->magenta

Gap table (ours vs best-of-breed)

axisoursbestholder
uv_textures0 (our exporter)1Seed-san + CharacterGen oracle (nx_modelfacts 2026-08-04)
vrm_emit0 (C3 rung open)1the entire VRM ecosystem; Seed-san = our round-trip oracle
proportion-edit-with-reexportrung filed 1785877444EMPTY CELL - no occupant found186-app showcase + EN/JP/BOOTH adversarial search 2026-08-04
calfLen_retargetCONVERGED 1.0002 (was INEXPRESSIBLE in the 9-joint export)1.13327 (theirs, calf/thigh)11-joint ankle-landmark export 2026-08-05; deformation rides the rig ruler 1785878464
form_headline_permil278 (atlas T2 being vs her clean body - highest measured, WITH a named ruler gap: it has no arms)1000 (her own mesh, same ruler)nx_bodybench full matrix 2026-08-05; detail axis binds all tube-bodies (175-232, naked-vs-dressed policy fork); next teeth: limb-presence judge + clothed-form geometry (1785935557)

Asset license (verbatim from the file's meta)

allowRedistributiontrue (read from the file's VRM 1.0 meta)
modificationallowModificationRedistribution (read from the file's meta)
creditREQUIRED - Seed-san (c) VirtualCast, Inc., licenseUrl https://vrm.dev/licenses/1.0/

Emitted by nx_asset_page from a judged manifest; hand-editing this page violates the publishing-house standing order. Sibling surfaces: /compare (per-domain axes) · /papers (research corpus).