pipelines/POSTMORTEM_LIKENESS_2026-08-08.md
The §8.14 FIRST CHARTER, executed and ruled. Charter: Josh graded the first textured turntable
(2026-08-07 night) "this is pathetic... why is this still so far off?... why are they all missing
our actual exemplary bodies and faces?" Five candidate root causes were adjudicated with evidence
by five read-only Opus lanes + a fresh completeness critic (workflow wf_e9327055-3a3, 1.16M
tokens, 307 tool uses, zero errors). Full evidence: the six annexes in
docs/pipeline_review/postmortem_likeness_annexes/. Every claim below carries its citation there.
The turntable Josh graded was **not the pipeline's best output — it was a stale, ungated,
self-disclaiming sighting render**: baked WITHOUT the head plates one hour before its own
head-plate tooling landed (verified: git show e045b2b2^ carries zero head_plates hits — a
timing accident, not a discipline failure), never de-lit (the MANDATORY stage 4b had never run),
wearing an untextured flat garment slot, with its head projected from ~15,000 source pixels onto
2.87M texels while a 506×1344 face plate sat unused on disk — and it reached Josh's eyes carrying
two RED lens verdicts that nothing consumed. Meanwhile the head itself is a bare cage wrap
(no ears, no hair, no eyes — each exclusion locally correct and ruled, their sum never scored by
any instrument), TRELLIS.2's native full-PBR surface was computed on this box twice and thrown
away at the export line both times, and no lens anywhere in the repo scores whole-figure likeness
to the plate. The system did not fail to iterate — it iterated brilliantly on measurable
sub-problems while nothing owned the whole read, and then published an artifact no gate had passed.
plate." The only whole-frame similarity instrument in the repo (copy_index,
face_overcopy_g4.py:74) lives exclusively in the 2D lane — nine files, all under
body_specs/, zero in build/3d — and was itself measured worse-than-useless for identity.
The re-frame the critic ratified: the chain does not degrade look SECRETLY — it trades look
for numbers EXPLICITLY, records each trade as a cost (ears −1440 verts, hair −29,760, the
sub-chin taper "price"), and has nothing that ever sums the costs. Of 16,288 measured points,
only ~21% enter any verdict; the 11,168 region-grid radii compare against nothing.
L01_proportion and L07_anatomical are GREEN on the asset Josh called pathetic. Magnitude of
per-stage look loss: UNMEASURED until the retrospective ladder runs (action C).
wrap its own graft record disclaims ("NO texture, NO hair, NO ears... NOT a photoreal pass",
GRAFT_A14_V3B_SKIRT_BALD.json:295). The profile "ear" is a photograph of an ear projected onto
a smooth dome (geometric z-slice scan: no lateral protrusion at ear height). The head texture
is mush by arithmetic: three full-body plates only (~15k source px on the head) onto 2.87M
texels ≈ 190×; bake_plates.py's own comment already diagnosed it; the head-plate path landed
one hour after the bake and the texture was never re-baked. CORRECTIONS the fix must carry:
the neck ring is a UDIM texture discontinuity, NOT a graft failure (0 non-manifold, rings
matched to 0.12cm) — a fix aimed at the graft would be aimed wrong; and the 2D stack is strong
only at the HEAD (body plates are 211×870 native upscales — a re-bake cannot fix the body).
The projection also painted the studio backdrop into skin albedo (34,676 texels), and R13_HAND
is over half invented (observed_fraction 0.49).
(every metrics file: bands/height_units/mesh/surface/symmetry/tag — no colour term exists),
and TRELLIS.2 ran twice AFTER its gate opened with the full PBR volume computed and discarded
at run_trellis2_a14.py:224-231 (only vertices+faces exported; no geometry-only mode exists in
the pipeline, so ~5GB of texture models were paid for in GPU time and dropped). The install is
complete and runnable today (16GB weights, nvdiffrast 0.4.0 compiled; blocker is the lane's own
import hook). One correction to §8.14's wording: TRELLIS.2 has NO multi-view API — "full-PBR
from the multi-view V3 plates" is a separate authored work item (T-MV); the native test runs
single-view FRONT first. The runnable FIRST TEST is authored in annex RC3 (action E).
a real model of the WRONG TOOL CLASS (video/world for physical AI; no character mesh; NVIDIA
routes character work to ARDY, which is motion). Edify 3D's NIM preview was retired 2025-06-06,
weights never released, partner-mediated stock APIs only, zero likeness evidence. On the field:
DreamCharacter-1's published character benchmark re-ranks the open tier with **TRELLIS-2.0 as
the strongest OPEN model on the character axis** (0.8329 ULIP, above Pixal3D and Hunyuan3D-2.1)
— independently supporting the RC3 test; Seed3D 1.0 (absent from the prior survey) is the top
new open candidate but its licence is UNVERIFIED at file level (BLOCKING before any download —
the §9 Hunyuan trap); the paid SaaS likeness leaders are Rodin Gen-2.5 (multi-view reference
conditioning + selectable quad topology) and Tripo P1/Prism 3.1 (best external facial-likeness
reads, n=1 sighted). HONEST COUNTERWEIGHT, unsoftened: not one paper or gallery in the sweep
shows a photoreal human face — a photoreal child face is unevidenced field-wide as of Aug 2026,
and no candidate is promised as closing that gap. RC4's proposed in-place supersession of
PIPE_CHARACTER_MODELS_2026-08-06 §3.1/§3.2 is HELD pending a second reader verifying the
benchmark table at source (its central claims are web-sourced with zero file:line citations).
L02 head-band exclusion is emitted with no expiry, no owner, no arming condition — a head with
no ears cannot fail any lens in the battery. The design (annex RC5, the build spec): four new
lenses beside the existing eight — L09 head-band-scored silhouette against a per-stage
HEAD-STATE declaration, L10 whole-figure perceptual (4 channels published separately, never
averaged), L11 face identity via SFace/YuNet (already installed in the pinned face312 venv;
licence line read-at-fetch owed), L12 ship-scale texture read that REFUSES pre-de-light — with
thresholds DERIVED from four corpora (plate-twin / different-subject / stage-delta /
planted-defect), the class-curve law binding, n<8 = descriptive-cannot-gate, and attribution =
first stage crossing its band. Two blockers before it can arm: the C1 pair counts are
uncounted, and ship-scale on-screen pixel height is undefined anywhere (action G).
candidates are about how the asset was MADE; the grade was produced by how it was SHOWN. Five
defects were knowable before Josh saw the frame. RULED as a RULE, not a new gate (creation-first
law): **nothing reaches Josh without (i) its own record's HONEST_TIER read aloud in the message,
(ii) a current eval of the exact artifact shown, (iii) the declared not-armed lenses listed.**
Binding on the director and every lane from this commit forward.
1. nvdiffrast evaluation-tier read — RULED PERMITTED. The licence's own §3.3 defines
non-commercial as "research or evaluation purposes only"; a route-deciding bake-off is
evaluation; the prohibition triggers on a shipped artifact. Per the standing face-value
licence law this is not a conservative-read case. Quarantine enforced in code (no
evaluation-tier output on any ship path); the cumesh/cubvh residual gates ship-tier only.
2. The stale census BANNED line — NARROWED IN PLACE (should_exist_census.py:304), as
MODEL_STACK_AUDIT_AUG2026 L345-348 already ordered. A census line reading BANNED was treated
as authority and kept the native texture unshown — the mechanism is itself a post-mortem
finding.
3. Sequencing — RULED (the critic's reconciliation adopted). Non-destructive zero-token work
first; regeneration only after the retrospective ladder is banked (the current artifacts are
the only pre-change ladder in existence):
A re-bake with head plates (CPU — RUNNING at this writing; positive control: the new
record's plates[] must carry kind/used_for keys) → B delight_atlases.py (mandatory §5) →
D re-run thousand_point_eval on the TEXTURED asset (L04/L05 fire for the first time ever) →
C the retrospective stage ladder, material-blind, one rig, every stage mesh, both arms,
BEFORE any regeneration → E the TRELLIS.2 native-PBR FIRST TEST (GPU ~15 min; validity
gate: seed-42 geometry must reproduce the recorded mesh) → F Seed3D 1.0 licence pin at
file level BEFORE any download → G count C1 pairs + define ship-scale pixel height.
4. Owed and boarded with owners: ear restoration (Josh already ruled "Agreed on ear pick";
NAMED OWED in CAGE_SOURCE_RECORD.json; mechanism = director call at build time), hair + eyes
stages (geometry exists — 29,760 hair verts/1,488 cards, corneas — no owner until now: the
likeness build wave owns them), ARMS_WHEN + OWNER on the head-band exclusion, the
asset_budgets surface decision for build/3d/finish_a14/**, the body-plate resolution ceiling
(the one thing no re-bake fixes), and the T-MV multi-view sampler authoring.
**FORK 1 — RC4 "Move 2": bench the paid SaaS likeness leaders on our own plates (~$20-40 +
an IP-out).** Buy one generation each from Rodin Gen-2.5 and Tripo P1 on the CHAR_0001 V3 plates.
FOR: Rodin is the best architectural fit in the sweep (multi-view conditioning — no open model
has it — and quad output inside our 24k budget, deleting retopo); Tripo/Prism carry the only
independent best-on-facial-likeness reads; an afternoon buys a hard upper bound on what the field
can do with our exact plates — and if the paid leaders ALSO miss, the post-mortem reframes off
tooling and onto our head/hair/skin stages, which is worth more than a model swap. AGAINST (the
lane's own strongest objection, carried verbatim): chasing a better generator may treat the wrong
root cause — a better mesh feeds the same assembly; and it uploads unreleased exemplar plates of
our protagonist to third-party SaaS under their retention terms. RECOMMENDATION: YES, run it in
parallel AFTER E gives the native-TRELLIS.2 read, as a discriminating experiment, not a
procurement — but the spend and the plate upload are Josh's to approve, not delegated.
STRONGEST OBJECTION to the recommendation: the IP-out is irreversible in a way the $40 is not.
FORK 2 — the ear/hair/eyes completion order vs another regeneration round. The evidence says
finish the head's missing CONTENT (ears ruled, hair/eyes exist as geometry) and re-show before
judging any generator swap — the current grade was rendered on an artifact missing all three plus
its own head texture. RECOMMENDATION: complete-and-reshow first (actions A-D land most of it this
sitting at zero token cost); regeneration decisions wait for the ladder + E. Objection: if the
ladder shows the generator sculpt was already off at S1, completion polish is spent on a wrong
base — which is exactly what the ladder exists to decide.
Action A running; B-G queued this sitting (B, D immediately after A; C's Blender batch and E's
reviewed script next; F/G five-minute jobs). The likeness-lens BUILD (RC5's W0-W1) is the
fresh-week wave's first charter, with the annex as its spec. RC4's held supersession goes to a
second-reader lane in the same wave. Consumers: docs/DOC_MAP.md row (this sitting),
HANDOFF §8.15 (the claiming seat), the likeness build wave's charter, and
build/ops/CAPACITY_ROTATION.md's Sat-1am queue (items 1-3 discharged by this document and its
actions; the RESIDUE LANDING WAVE remains item 4).
plates (3 body + bust + face routed to the head tile) with kind/used_for keys, zero REFUSED.
A14_BASECOLOR_DELIT.{1001,1002,1003}.png) — stage 4b'sfirst-ever run.
because the assembled OBJ's material library declared no map_Kd — the graded turntable had
bound its atlases by Blender slot-table only, so the asset never referenced its own textures.
Re-assembled through the round-2 concatenator (the map_Kd fix), re-evaluated:
**L04_tone GREEN and L05_colour AMBER (worst R05_PELVIC_GIRDLE = the garment area) — the two
lenses scored for the first time in the harness's existence.** Card at
build/3d/eval/thousand_point/EVAL_A14_ASSET_TEX_V4.json; the refused intermediate card is
kept as EVAL_A14_ASSET_TEX_HEADPLATES.json. The geometry REDs (L02/L03) stand, as expected.
191,732 textured faces, real 2048² basecolor + metallic-roughness, UVs, webp. **VALIDITY
CONTROL PASSED: seed-42 geometry reproduces the recorded run exactly (940,424 v / 1,895,540 f)
— texture is the only variable.** One trap caught by sighting: the GLB imports 180° off the
house frame (the extent assert cannot catch a flip); the turntable's az000 is the back, and
the sheet remaps +180 with the remap printed on its face.
build/3d/bakeoff/native_tex/SHEET_NATIVE_VS_CHAIN_A14.png — plate / gradedV3 / re-baked V4 / native, four views each. **THE EYE'S READ (director sighting, to be
confirmed by Josh):** the native row is the only 3D row that reads as a CHARACTER — pointed
elven ears, the plate's black trunks, coherent face and unified skin — while both chain rows
keep the collapsed shard shoulders, the partition-border scars printing through the face, and
the untextured cream trousers. **RC3's inversion condition is met on sight: the architecture
inverts to generate-rich / fit-lightly / preserve-the-generator's-surface.** Honest
counterweight, carried with it: the native face reads ADULT — it is not yet the 14-year-old
boy — so native wins whole-read coherence while identity-to-the-plate remains open; that is
exactly the likeness lens's L11 question, and the T-MV multi-view arm + conditioning work is
where the age likeness gets fought.
build wave (wf_7ced0644-9c9) is running: the L09-L12 build + C1 pair counts + the RC4
second-reader + the Seed3D licence pin.
and the benchmark is ANIME-DOMAIN (PAniC-3D/AnimeRecon, ULIP/Uni3D embeddings, VRoid/VTuber
corpus), so "TRELLIS-2.0 strongest open on the character axis" carries that qualifier
everywhere it is repeated, and PIPE_CHARACTER_MODELS §3.1's class finding SURVIVES: a
photoreal human face is unevidenced field-wide. The supersession landed in place
(§3.1 append + three §3.2 cell edits + the Seed3D note), per the reader's exact wording.
"license": null, no weights under either HF org (controlled absence), API-only on Volcano.
Action F closes PERMANENTLY-BY-FACT; evidence pinned at docs/licence_records/seed3d/ with
the 12-URL attempt log. This section of the RC4 annex (§7 item 1, §6's local-fit row, §0
headline 2) is REFUTED and must not be cited; RC4's "reusable benchmark protocol" (VLM
camera rule etc.) has NO located source and must not be copied.
RC4's table elided for exactly its two recommended tools — Rodin 2.5 "facial details were
noticeably softer" + "Weak texture quality"; Trellis 2 "more sculpted than realistic". So if
Josh approves the spend: **Tripo P1 / Prism 3.1 is the likeness arm the evidence supports;
Rodin is the ARCHITECTURE bet** (multi-view conditioning + quad topology), not the likeness
bet, and is presented to him as such.
lens) is ARMED and mutation-proven, the crop-view refusal is textbook, and action G's first
half is discharged with numbers (C1 strict pairs: face=3, every full-body view=0 — the
absolute leg is DESCRIPTIVE-CANNOT-GATE until second renders exist). But the L10 self-test's
"clean twin" was byte-identical to its plate, collapsing every must-fire bar to an epsilon
guard (mutation-proven: a 1e-6 sensitivity collapse survives the self-test), and against a
REAL regeneration negative MS-SSIM and GMSD INVERT — the copy_index defect class itself.
Plus: L12's de-light gate reads a key no record emits (and its "never ran" premise is false
on disk since 01:26), the garment exclusion is claimed in the docstring and not implemented,
and the attribution half is unarmed. Verdict + the nine-fix worklist:
build/3d/eval/likeness/CRITIC_VERDICT_W0.json. **THE FIX PASS is chartered on that
worklist; keep-as-is list honored (L09, crop refusal, C1 counts, corpus verdict).**