The hospital gown renders as shards on two stations, and it is the only garment 4-15x the fleet's triangle count

[gidich]

The hospital gown's surface renders as triangular shards wherever it appears — in the live ED room and
in an isolated bake capture of a different actor. It is also the only garment in the fleet whose
triangle count is an order of magnitude out of line.

MY GRADE, native resolution, no upscale

tools/openclinxr/evidence/issue-714/bodice-band-1to1.png, 1060x585, a 1:1 crop of a 4096 render of
mpfb-gown-adult-patient.glb through the model-vetting three.js path:

• triangular sawtooth spikes running down both vertical seam lines
• a shattered region at the sternum and upper chest
• spikes along the shoulder and the lower hem
• two orange-tan slivers mid-chest and one upper-left that are skin visible through the garment, so
  that region is open rather than merely faceted

#712's measurement found the same appearance in the live ed_chest_pain_priority_v1 capture, and
established what it is NOT: not skinning weights, not a bone transform, not a morph, not vertex
displacement. The mesh's current AABB sits within 1.25x of its bind AABB and the pixels land exactly
where the measured geometry is.

Measured 2026-08-28 across every shipped humanoid

asset	gown/garment primitive	triangles	vertices
mpfb-gown-inspect	..._hospital_gown_phenotype_L0	57,073	28,852
mpfb-gown-adult-patient	..._hospital_gown_phenotype_L0	29,185	14,899
ed_chest_pain_spouse_adult	open cardigan L1	7,829	4,468
ed_chest_pain_spouse_adult	casual top L0	4,146	2,161
peds_anxious_parent	open cardigan L1	8,009	4,506
peds_anxious_parent	casual top L0	4,069	2,101
peds_fever_patient_child	phenotype L0	4,883	2,565
peds_nurse_kevin	scrub pocket L1	4,076	2,106
peds_nurse_kevin	scrub top L0	4,170	2,158
peds_patient_child	exam t-shirt L0	3,887	2,042

Every non-gown real garment falls between 3,887 and 8,009 triangles. The hospital gown carries 4x to
15x that, and it is the only one that renders as shards.

What is NOT established

That the triangle density is the cause. It is a correlation across nine primitives in seven
assets, and it is equally consistent with density being a symptom of the same fold generation that
produces the shards. CAUSE NOT DETERMINED — do not take the correlation as a mechanism.

The two obvious readings, unranked, and they may both be wrong: the fold wave
(_fold_amp686 = 0.034, _fold_k686 = 16 in automate_blender.py) subdivides the surface past what
the shell can carry; or the gown's shell is generated at a different base resolution from every other
garment and the wave merely makes the result visible.

The first measurement, before any product edit

Compare the hospital gown's surface against a garment that renders clean — peds_nurse_kevin's scrub
top at 4,170 triangles is the known-good column, same pipeline, same actor family. Report edge-length
distribution, triangle aspect ratios, normal discontinuity across adjacent faces, and boundary-edge
count for both. That says whether the gown's surface is degenerate or merely dense.

tools/openclinxr/evidence/gown-shard-mechanism-measure.ts already walks the gown asset and is the
instrument to extend rather than replace.

Do not

• Do not lower _fold_amp686 or _fold_k686. The bodice fold amplitude is 34 mm on a 10-22 mm standoff, so the wave valleys land inside the body #714's clause (3) and 73 gown vertices are inside the body at y 1.153-1.348 m after #714's derived clamp #746's counterweight (2) both
  refuse it, and flattening the gathers is not a fix for a surface that is shattered.
• Do not decimate the gown to bring the triangle count into line. That would move the number this card
  reports without establishing what the shards are, and Wire meshopt decimation into the humanoid pipeline — the silhouette holds at 2.81x and the station comes under budget #695 owns decimation.

Report these

If any proof cannot pass as written, OR passes trivially against the ambient measured range, OR
asserts the opposite direction from the defect, say so in your first report, at the moment you find
it.

If the premise here turns out to be false, say so immediately. Never narrow or weaken a clause so that
it passes.

Report any out-of-scope wrongness you saw and are not fixing, naming the object and what is wrong. The
detached trouser waistband (#320) and the unlit room (#526) are already carded.

Write your report with the section headers IN-SCOPE:, OUT-OF-SCOPE:, CLAIM: and NOT TESTED: in
that colon form.

factory_step: clothing_consume

The RED is planted at 006bac1a

tools/openclinxr/evidence/the-hospital-gown-surface-is-as-smooth-as-a-garment-that-renders.test.ts.
Clause (1) is it.fails and reds on the real defect; three counterweights are green.

The metric is the share of interior edges whose two faces meet past 60 degrees, from GEOMETRIC face
normals derived from positions, after welding vertices by position at 5 decimal places. It is immune
to shading and to index splits at material or UV seams, and it is what "a field of triangular spikes"
means numerically. It deliberately does NOT bound the triangle count — a count is a quantity and
shards are a shape, so a count clause greens the moment the mesh is decimated while the surface stays
shredded.

Clause (1)'s threshold is measured at test time from the non-gown population rather than chosen here.
It resolves today to 3.47% against the gown's 19.32%, a 5.6x margin.

Destructive probes, all three fired and all three reverted, recorded in the plant commit:
pointing both gown subjects at clean geometry makes clause (1) pass; tightening the comparator ceiling
to 2.0% reds counterweight (3); lowering _fold_amp686 to 0.010 reds counterweight (2).

If the surface cannot be fixed inside this slice

Report the measurement and stop. That closes this card successfully.

The live: rule requires clause (1) to be flipped, and a clause you cannot honestly flip will fail
it. That is a defect in my contract, not in your work, and I will resolve it the way #714 was: the
unmet clause is relocated verbatim to a new card with an inverted guard at the origin. Do not flip
the clause to make the proof pass, and do not weaken it.

CORRECTED 2026-08-28 (second pass) — read this before the sections above

Three candidate causes are measured dead. Do not re-derive them.

candidate	verdict	evidence
the C0 wave corner	contributes 0.65 of 16 points	a phase-matched C1 cosine at the same amplitude and wave count moved the adult gown 19.32% → 18.67% and inspect 21.08% → 20.70%; gathers survived exactly
aliasing	REFUTED	two horizontal rings carry 315 and 197 distinct azimuth columns with phase concentration R = 0.067 and 0.034 — ~19 samples per period. torso_cols = 16 is the parametric grid, NOT the exported azimuthal count
slivers	REFUTED	sharp fraction against a minimum triangle area of 0 / 1e-8 / 1e-6 / 1e-5 m²: gown 19.32 / 18.01 / 16.98 / 11.58%, scrub top flat at 1.29%. At 1e-6 the gown is still 13x

The C1 attempt is preserved at branch salvage/issue-747-c1-wave (bb461d64). It is NOT landed: it
changes shipped GLB bytes for 0.65 points, changes nothing a grader can see, and raises the
near-zero-area triangle count from 483 to 2,636 on the adult gown. Read it if useful; do not assume it.

What survives, as a HYPOTHESIS and not a measurement: the wave is too deep for its wavelength. At
k = 16 on a ~2.2 m chest circumference a half period is ~69 mm, and a 34 mm excursion gives a crest
dihedral near 90 degrees — which is the measured p90 of 86–90. CAUSE STILL NOT DETERMINED. The
arithmetic fits; nothing has tested it.

Counterweight (2) was amended at 6ffa5977 because it bounded the fix out

It pinned _fold_amp686 = 0.034 AND _fold_k686 = 16 as literal equalities, so every remaining
candidate was refused by the clause meant to stop one specific cheat. That was my defect.

It now bounds the property those constants stood in for: the gather's DEPTH, as an amplitude FLOOR at
0.030 — 88% of the shipped 0.034, so a trim is allowed and a flattening is not. The wave count is
unpinned on purpose: lowering k lengthens the wavelength and reduces the crest slope while leaving
depth untouched, which is the opposite of flattening.

Probes, all three fired and reverted: _fold_amp686 = 0.010 reds it; _fold_k686 = 8 stays green, so
the wavelength fix is legal; _fold_amp686 = 0.029 reds it.

The work is a SWEEP, and you do not pick the value

Do not choose an amplitude and a wave count and defend them. Produce a variant sweep and let the
orchestrator grade it.

Sweep _fold_k686 across at least four values spanning a real range (something like 4, 6, 8, 12) and
_fold_amp686 across at least two at or above the 0.030 floor. For each cell, rebake
mpfb-gown-adult-patient from current main and record the sharp fraction, the near-zero-area triangle
count, and the peak radial excursion into gown-vs-clean-surface.json beside the existing rows.

Then produce ONE labelled contact sheet of the rendered bodice band across the cells, native
resolution, on a TRACKED path under tools/openclinxr/evidence/issue-747/. Label every cell with its
k and amplitude. The orchestrator grades the sheet and picks the cell; you do not need clause (1) to
go green to succeed.

If no cell clears the ceiling, that is a finding and it closes the slice. Report the table and
stop. Do not chase the threshold, do not go below the amplitude floor, and do not decimate — #695 owns
decimation and clause (1) bounds a shape precisely so decimation cannot green it.

Duration is not a constraint here. A multi-hour bake sweep is acceptable and preferred over a guess.

done_when

• run:pnpm exec vitest run tools/openclinxr/evidence/the-hospital-gown-surface-is-as-smooth-as-a-garment-that-renders.test.ts
• live:tools/openclinxr/evidence/the-hospital-gown-surface-is-as-smooth-as-a-garment-that-renders.test.ts
• exists:tools/openclinxr/evidence/issue-747/gown-vs-clean-surface.json

gown-vs-clean-surface.json is the pre-fix comparison, written on a TRACKED path BEFORE any product
edit — a deliverable under a gitignored path has no land path. One row per garment primitive carrying
edge-length distribution, triangle aspect ratio, the dihedral distribution, and boundary-edge count,
so the question "is the gown's surface degenerate or merely dense" is answered from data rather than
from the triangle count.

[gidich]

Factory: Dispatched — measurement landed; clause (1) NOT flipped (surface not fixable inside this slice within the card's constraints; the card's own report-and-stop path applies).

IN-SCOPE:

• Pre-fix measurement written on a TRACKED path BEFORE any product edit: tools/openclinxr/evidence/issue-747/gown-vs-clean-surface.json (commits f7e92c69, a6161eb6), produced by extending gown-shard-mechanism-measure.ts (env-gated OPENCLINXR_GOWN_747_SURFACE=1 mode, same seven primitives as the planted contract).
• The card's DEGENERATE-or-DENSE question, answered from data: the gown is DENSE, not degenerate. Edge p50 9-12mm (clean 18-23mm), aspect p90 4.1-4.4 (clean 2.2-2.5; slivers >10 are 0.7-0.8% of triangles), degenerate-area triangles 1.7-3.7%. The discriminator is the DIHEDRAL: gown p90 86-90 deg, frac>60 19.32-21.08%; clean p90 17-33 deg, frac>60 1.3-3.5%.
• Natural control answered: the two gown primitives differ by 2x in triangles (57,073 vs 29,185) yet sit within 2 points on the sharp fraction (21.08 vs 19.32) — density is NOT the mechanism. Reading (2) of the card is FALSE: the gown's base resolution (torso_rows/cols 11x16, automate_blender.py:3996) equals the open cardigan's (11x16, :4010), which renders clean at 3.47%.
• Mechanism named: the The hospital gown's bodice is shrink-wrapped — anatomy reads through the fabric above the waist while the skirt drapes #686 fold block (automate_blender.py:2562-2671, gown-only — skirt_top_y is set only for the gown kind) subdivides the shell 2x (that is the 4-15x triangle count) and displaces the bodice by a C0 TRIANGLE wave (_tri_wave686, amp 0.034, k=16). The 32 vertical crest/valley fold lines are the hard >60 deg dihedral edges — the "triangular sawtooth spikes running down both vertical seam lines" of the grade. Sharp edges localize to the fold band (bands 4-9 = 15-32%, peak 29.6%) vs the skirt (10-13%); boundary edges are mid-bodice (540/843 on the adult in bands 4-8, peak 222 at y 1.153-1.250 = sternum), confirming the grade's "open rather than merely faceted" from welded-edge data, not pixels.
• Fix identified but not landable inside this slice: replace the C0 triangle wave with a C1 smooth wave at the SAME amplitude and count (keeps _fold_amp686 = 0.034 / _fold_k686 = 16 — counterweight (2) and The bodice fold amplitude is 34 mm on a 10-22 mm standoff, so the wave valleys land inside the body #714/73 gown vertices are inside the body at y 1.153-1.348 m after #714's derived clamp #746 regexes pass; keeps the gathers; removes the hard corners). Landing it requires re-baking both shipped GLBs at apps/ui-xr/public/generated-humanoids/ — outside the asset-pipeline write roots — and the The bodice fold amplitude is 34 mm on a 10-22 mm standoff, so the wave valleys land inside the body #714-style rebake splice used meshopt decimation, which this card forbids. Per the card: "report the measurement and stop."

OUT-OF-SCOPE:

• Pre-existing reds at the plant commit (not caused by this change): the-gown-shards-are-measured-before-anything-is-changed.test.ts and the-gown-shell-clearance-is-measured-not-assumed.test.ts pin the adult-patient GLB at 7,116,988 B / sha 7bd12d06; the tree's asset is 9,280,276 B / sha dd2340a9 (moved by the The bodice fold amplitude is 34 mm on a 10-22 mm standoff, so the wave valleys land inside the body #714 rebake). Pins not mine to update.
• pnpm architecture gate fails on markdown-references (PROJECT_STATUS.md: 4 unresolved refs; doc-authority-registry .md: 1) — pre-existing, unrelated; this change adds no markdown.
• No product code or asset bytes changed. Text-only worker: no pixel grading; the grade's sliver/tear appearances were checked from geometry only.

CLAIM:

• Demonstrated from welded-geometry data that the gown's surface is DENSE not degenerate, that density is not the shard mechanism (2x-triangle control), and that the shards are the C0 triangle-wave fold lines of the The hospital gown's bodice is shrink-wrapped — anatomy reads through the fabric above the waist while the skirt drapes #686 block (localized to the fold band, open mid-bodice edges at sternum height) — the fix is a smooth wave at unchanged amplitude/count, which cannot be landed inside this slice's write scope and constraints.

NOT TESTED:

• Whether a sine wave at amp 0.034 / k=16 actually clears the 3.47% ceiling (needs a control bake + measurement; the bake path writes outside this role's write roots). Whether the gathers visually survive any wave change (orchestrator render grade required). Whether the boundary edges are all fold-induced vs. partly export/weld artifacts.

[gidich]

Worker returned at 68 turns, clean worktree, 2 commits, ready_to_integrate. Three of four proofs
pass; live: fails because clause (1) stays red, which is the stop the card pre-authorised. No asset
bytes changed, no product code touched, and the contract file is untouched — one deletion across the
whole branch, a refactored const.

I was wrong about "degenerate", and the correction is the worker's

Reading the artifact mid-run I said the 483 and 2,124 near-zero-area triangles read as DEGENERATE
rather than dense. That was the wrong comparison: as a fraction of triangles the slivers are 0.7–0.8%,
and aspect p90 is 4.1–4.4 against 2.2–2.5 on clean garments — a difference in degree, not in kind. The
discriminator is the dihedral, at 19–21% against 1.3–3.5%. The gown is dense, not degenerate.

The residual I still cannot explain: every clean garment carries 0 or 1 near-zero-area triangles and
the gown carries 483 and 2,124. That is a real difference the "dense" reading does not account for.

Reading (2) of this card is refuted, and I verified it myself

I wrote that the gown's shell might be generated at a different base resolution. It is not.
automate_blender.py sets torso_rows, torso_cols = 11, 16 for the gown and 11, 16 for the open
cardigan, which renders clean at 3.47%. Same resolution, 5.6x apart on the metric.

The mechanism, verified at source

_tri_wave686 at automate_blender.py:2574 is a piecewise-linear triangle wave: it returns
1 - 2u/pi on the first half period and -1 + 2(u-pi)/pi on the second. Its value is continuous and
its derivative is not — every crest and every valley is a slope reversal, so the surface carries a
hard ridge line there by construction. At _fold_k686 = 16 that is 16 crests and 16 valleys running
vertically down the bodice, strongest at the centreline where wx peaks.

That matches my pixel grade exactly: triangular sawtooth spikes running down both vertical seam lines,
worst at the sternum.

And the openness has the same locus

540 of the adult gown's 843 boundary edges sit in bodice bands 4–8 (y 0.958–1.445), peaking at 222 in
band 6 at sternum height. Clean garments' boundary edges are rim- and opening-concentrated. So the
open region I graded at the sternum is not a rim: it is where the fold wave folds the surface back on
itself.

The natural control holds

57,073 triangles against 29,185 — 2x — while the sharp fraction sits within 2 points, 21.08% against
19.32%. Density is not the mechanism.

CLAIM: the gown's shards are the C0 discontinuity of _tri_wave686, not mesh density and not base
resolution. Verified by me at source: the wave is piecewise-linear, and the gown shares 11, 16 with
a garment that renders clean.

NOT TESTED: whether a C1 wave of the same amplitude and wave count removes the ridges without
flattening the gathers. That is the fix and it is the resume.

[gidich]

Factory: Dispatched — fix attempted and measured; clause (1) NOT flipped (post-fix worst gown 20.70% vs ceiling 3.47%; per the operator's constraint 4, report the number and stop).

IN-SCOPE:

• C1 wave built and landed: _tri_wave686 -> _fold_wave686 = cos(x + pi/2) = -sin(x) in automate_blender.py:2574-2588, same amplitude (0.034) and same wave count (16); _fold_amp686 = 0.034 / _fold_k686 = 16 literally untouched (all counterweight regexes pass). Raised cosine 0.5*(1+cos(x)) rejected: +0.5 DC offset shifts the whole fold band outward and kills the inward valleys.
• Rebake (constraint 3): both gown GLBs re-baked from current main via bake_mpfb_gown_inspect.py + splice-gown-747.mjs (gown-only replacement, JOINTS_0 remapped by name into the shipped 138-joint skin; weight sums all 1.0; 9 non-gown meshes byte-identical; 17 materials before/after, embedded base-colour textures intact on body/shoes/t-shirt/eyes; no decimation).
• Post-fix sharp fraction (constraint 4), welded 5dp geometric dihedral, the contract's metric: adult 19.32% -> 18.67%; inspect 21.08% -> 20.70%. Ceiling (non-gown population) = 3.47%. IMPROVES BUT DOES NOT CLEAR — reported, stopped, clause (1) stays red, not flipped, not weakened. The live: proof fails for that reason, as the card anticipates (clause relocation).
• Gathers survive (constraint 2): fold-band peak radial excursion adult max 0.4912 pre and post; inspect max 0.4578 pre and post. The gathers did not vanish.
• Near-zero-area triangles (the second-mechanism question): NOT fixed by the C1 wave. pre-fix adult 483 -> post-fix 2636; inspect 2124 -> 2603 (clean garments carry 0-1). This population is a SECOND mechanism, not the wave corners.
• Fresh native 4096 render + 1:1 bodice-band crop staged on tracked path tools/openclinxr/evidence/issue-747/ (front/three-quarter lit+structure + bodice-band-1to1.png, fold-band rows 1273..2131 by camera projection). Orchestrator grades; text-only worker does not.
• gown-vs-clean-surface.json updated: pre-fix and post-fix rows in one artifact (14 rows, 7+7).

OUT-OF-SCOPE:

• Pre-existing reds, verified by stash-testing (fail identically without my GLB change): the-gown-mhclo-fits-the-stripped-basemesh (reads .openclinxr-local provider-cache paths absent in this worktree); garments-meet-at-the-waist (mpfb-clinical-nurse-adult, not a gown asset); the-gown-shards-are-measured and shell-clearance (byte pins predating The bodice fold amplitude is 34 mm on a 10-22 mm standoff, so the wave valleys land inside the body #714); pnpm architecture (markdown-references).
• The shard mechanism is NOT (only) the C0 wave corners: a C1 wave at the same amplitude/count leaves frac>60 essentially unchanged. The measured shards sit in the fold band AND the skirt bands; the degenerate-triangle population is independent of the wave shape.

CLAIM:

• Demonstrated, with a measured rebake of both shipped gown GLBs, that replacing the C0 triangle wave with a C1 cosine at unchanged amplitude (0.034) and count (16) preserves the gathers (peak radial excursion unchanged) and textures/skin binding, but moves the sharp-edge fraction only 19.32->18.67% and 21.08->20.70% — far above the 3.47% ceiling, so clause (1) cannot honestly flip.

NOT TESTED:

• Why the C1 wave left frac>60 essentially unchanged (the remaining shard mechanism is not the wave corners; the fold-band and skirt sharpness and the 483-2636 degenerate triangles are still unexplained). Whether a different C1 shape or any other fix clears 3.47%. Orchestrator render grade of tools/openclinxr/evidence/issue-747/*.png.

[gidich]

The C1 wave is not the fix, and my mechanism claim was wrong as a CAUSE. Preserved at
salvage/issue-747-c1-wave (bb461d64); not landed.

What the treatment did

primitive	pre-fix	post-C1	ceiling
hospital gown (adult)	19.32%	18.67%	3.47%
hospital gown (inspect)	21.08%	20.70%	3.47%

Gathers survived exactly — peak radial excursion 0.4912 adult and 0.4578 inspect, identical before and
after. Amplitude and wave count stayed literal. And the near-zero-area triangle count got worse:
483 → 2,636 on the adult, 2,124 → 2,603 on inspect.

MY GRADE of the post-fix render

tools/openclinxr/evidence/issue-747/bodice-band-1to1.png, 1060x858 native, no upscale.

Still shredded. Dense triangular spikes down the centre-chest columns, a torn region from the sternum
to the collar with skin visible through it in three places, spikes along both flanks, tan rectangles
at both upper arms. The shoulders and outer chest do read as smooth cloth now and the spikes are more
concentrated toward the centreline, which matches a 0.65-point improvement and nothing more.

Correcting myself

I verified at source that _tri_wave686 is piecewise-linear and stated the C0 discontinuity as the
mechanism. The property is real; it is not what produces the 60-degree edges, because removing it
moved the metric by 0.65 points out of a required 16.

The reading that fits every observation, and it is a HYPOTHESIS

_fold_k686 = 16 and the gown's torso_cols = 16. The fold argument is
_fold_k686 * atan2(rz, rx), so consecutive columns, spaced about 2pi/16 apart in azimuth, step the
wave argument by about 2pi — every column samples the wave at nearly the same phase. The wave is
aliased. What reaches the mesh is not sixteen gathers but the residue of irregular column spacing
against a function varying far faster than the sampling, which is noise, and noise is what the pixels
show: not sixteen regular folds but a spray of spikes.

It explains the C1 null result — an unrepresentable wave is unrepresentable whatever its smoothness —
and it explains the open cardigan, which shares torso_cols = 16 and renders clean at 3.47% because
the fold wave is gown-only (OPENCLINXR_SUPPRESS_GOWN_FOLD686, automate_blender.py:2598).

NOT MEASURED. I have not sampled the actual per-column phase. The test is cheap: dump the fold
argument per column and see whether the phases collapse.

My own counterweight forbids the likely fix

Counterweight (2) pins _fold_k686 = 16 literally. If aliasing is the cause, the fix is to make the
wave representable — lower k, or raise torso_cols — and the first of those is banned by a clause I
wrote to stop the gathers being flattened.

That is the wrong bound for the job. Flattening is an AMPLITUDE move; wave count is a different axis.
The counterweight should pin _fold_amp686 and require the wave to remain representable —
torso_cols >= 2 * _fold_k686 — so k may fall or cols may rise while the gathers cannot be
removed. I will amend it before the next slice rather than ask a worker to work around it.

CLAIM: replacing the C0 triangle wave with a phase-matched C1 cosine at the same amplitude and wave
count improves the sharp-edge fraction by 0.65 and 0.38 points, leaves the gathers unchanged, worsens
the near-zero-area triangle count, and does not change what the bodice looks like.

NOT TESTED: the aliasing hypothesis itself. Whether the near-zero-area triangles are a second
mechanism or a consequence of the same sampling. And whether a representable wave produces gathers a
grader would accept, which only a render can answer.

[gidich]

I tested my own aliasing hypothesis and it is refuted. Two more candidates fall with it, and what
survives is a bound I wrote into this card's own counterweight.

Aliasing: REFUTED, measured

Sampling the shipped gown at two horizontal rings in the fold band and computing the wave phase
(16 * atan2(rz, rx)) mod 2*pi per distinct azimuth column:

ring	distinct columns	phase concentration R
y 1.29–1.31	315	0.067
y 1.19–1.21	197	0.034

R near 0 is a uniform phase spread; R near 1 would be the collapse aliasing predicts. The gown's
bodice carries roughly 300 azimuth columns per ring, not the 16 that torso_cols suggested, so the
wave is sampled about 19 times per period — comfortably representable. torso_cols = 16 is the
parametric grid, not the azimuthal count of the exported mesh, and I read one as the other.

Slivers: also not the cause

Recomputing the sharp fraction while discarding any edge whose either face is below an area floor:

min triangle area	gown adult	scrub top
0 (all)	19.32%	1.29%
1e-8 m²	18.01%	1.29%
1e-6 m²	16.98%	1.29%
1e-5 m²	11.58%	1.29%

At 1e-6 the gown is still 13x the scrub top. The 1e-5 row has discarded a quarter of the gown's edges
and is over-filtered. The sharp edges sit on real, well-sized triangles.

What survives: the wave is too deep for its wavelength

Three candidates are gone — the C0 corner moved the metric 0.65 points, aliasing is refuted, slivers
are refuted — and the arithmetic on what is left is straightforward. Chest circumference is roughly
2*pi*0.35 ≈ 2.2 m, so at k = 16 one period is about 137 mm and a half period about 69 mm. Against
that, _fold_amp686 = 0.034 is a 34 mm excursion, giving a peak slope near atan(0.068/0.069) ≈ 44 degrees on each flank and therefore a dihedral approaching 90 degrees across every crest.

That is exactly the measured p90 of 86–90 degrees. The wave is not badly shaped and it is not badly
sampled; it is too deep for how often it repeats.

HYPOTHESIS, not measured. The test is a sweep of amplitude against wave count with the excursion
held visible, graded on renders rather than argued from trigonometry.

Counterweight (2) bounds the fix out, and that is my defect

It pins _fold_amp686 = 0.034 and _fold_k686 = 16 literally, to stop "flatten the gathers until
nothing is sharp". If the amplitude-to-wavelength ratio is the cause, every fix moves one of those two
numbers and my clause refuses all of them.

The bound is on the wrong quantity. What must not be lost is the VISIBLE GATHER, and its measure is
the peak radial excursion — 0.4912 m on the adult gown, unchanged by the C1 attempt. A clause pinning
that lets amplitude and wave count move together while still refusing a flattened bodice, and it is a
shape bound rather than a constant.

I will amend the counterweight before the next slice rather than ask a worker to work around it.

CLAIM: on the shipped gown the fold wave is sampled ~19 times per period and its sharp edges survive
sliver filtering, so neither aliasing nor degenerate triangles explains an 86–90 degree p90 dihedral;
the amplitude-to-wavelength ratio does, arithmetically.

NOT TESTED: the ratio hypothesis itself, and whether a shallower or longer wave produces gathers a
grader accepts. Also untested: whether the 2,636 near-zero-area triangles have their own cause, since
removing them from the metric changes it by only 1.3 points.

[gidich]

Factory: Landed (sweep deliverables; clause (1) NOT flipped — no cell cleared the ceiling)

IN-SCOPE:

The sweep is complete and decisive: NO cell clears the 3.47% ceiling (the open cardigan, measured at test time from the non-gown population). The best cell — k4 / amp 0.030 — sits at 15.46%, 4.5x above the ceiling. Reported per the card: the surface cannot be fixed inside this slice by the fold constants; clause (1) stays red (it.fails), not flipped, not weakened. The live: proof will fail for that reason — the anticipated outcome, resolved by relocating the unmet clause.

What was measured (each cell: full bake via bake_mpfb_gown_inspect.py with env-overridden fold constants at the use site, gown spliced into the SHIPPED mpfb-gown-adult-patient.glb — gown-only replacement, JOINTS_0 remapped by name, non-gown meshes untouched — so every row is measured in the shipped configuration):

cell	sharp >60deg	near-zero-area (<1e-8 m2)	peak radial (m)
k16 / a0.034 (control)	20.80%	2480	0.4575
k12 / a0.034	18.45%	2547	0.4575
k12 / a0.030	18.20%	2518	0.4575
k8 / a0.034	16.67%	2747	0.4575
k8 / a0.030	16.44%	2680	0.4575
k6 / a0.034	16.01%	2648	0.4575
k6 / a0.030	15.80%	2624	0.4575
k4 / a0.034	15.56%	2844	0.4575
k4 / a0.030	15.46%	2795	0.4575

Validation: the control cell (shipped constants) measures 20.80% on the spliced full-res gown, bracketed by the shipped decimated asset (19.32%) and the inspect asset (21.08%) — the instrument reproduces the family. The gathers survive at every cell (peak radial constant 0.4575; the fold only changes their count, not the crest reach).

The ratio reading (wave too deep for its wavelength) is PARTIALLY REFUTED. Lowering k helps monotonically (20.80% -> 15.46%) but saturates: the response halves each time k halves. Per-decile localization on the k4 cell shows the residual splits into (a) a k-invariant non-fold floor — skirt/waist bands (y 0.53-0.99, untouched by the fold) carry 11.7-16.6% sharp, and (b) the fold band itself still at 15-22%. Two further probes: the #714 trough clamp is NOT the floor (clamp-disabled bake at k4: 15.43% vs 15.56%); the sliver filter (card's second pass) leaves 11.58% at 1e-5. Notably, a clean garment's OWN hem band runs 10-13.5% sharp (normal rim creasing), but the gown's mid-skirt bands have no clean-garment counterpart — the lofted skirt construction itself carries creases no fold parameter touches. Zeroing the fold bands in the k4 decile data leaves ~5.0% — above the ceiling even with a perfectly smooth fold band.

Proofs and pre-existing reds:

• run: (the planted contract file) PASSES as planted: 3 passed + 1 expected fail. Clause (1) is it.fails and still red; counterweights (2)(3)(4) green, including (2) after the automate_blender edit (the literal _fold_amp686 = 0.034 assignment is untouched; the sweep overrides read at the use site, and an unset env reproduces the shipped bake).
• live: WILL FAIL — clause (1) is not flipped because no cell clears. That is the card's anticipated "surface cannot be fixed" path.
• exists: gown-vs-clean-surface.json — tracked, committed, 7 pre-fix rows + 10 sweep rows (8 cells + control + clamp probe).
• pnpm packages:typecheck:agent: 67/67 green. pnpm architecture: FAILS on the pre-existing markdown-references violation (PROJECT_STATUS.md 4 unresolved refs to gitignored .openclinxr/handoffs/* and wiki/index.md; doc-authority-registry 1 ref to studio dist/). Verified pre-existing: it fails identically with my changes stashed, and the C1 branch recorded the same red. Not caused by this slice.
• Two gown tests fail on stale byte pins predating this tree: the-gown-shards-are-measured-before-anything-is-changed and the-gown-shell-clearance-is-measured-not-assumed (GLB sha256/size pins). Verified pre-existing via stash. The other two gown tests importing the extended measure file pass.

Deliverables on tracked paths: tools/openclinxr/evidence/issue-747/gown-vs-clean-surface.json (pre-fix + sweep rows), bodice-band-contact-sheet.png (ONE labelled sheet, 9 tiles, 5010x2778, native 1:1 crops of the 4096 glb-grade front renders, each tile labelled k / amp; per-tile gown-blue pixel fraction 18.5-20.1% recorded as machine validation that the gown is in every tile), build-bodice-contact-sheet.ts (builder + crop provenance), splice-gown-747.mjs (gown-only splice tool), env-overridable fold constants in automate_blender.py.

OUT-OF-SCOPE:

• The shipped mpfb-gown-adult-patient.glb carries a mpfb_robert_reference body while the current regeneration input mpfb-viseme-inspect.glb carries an mpfb_ob_patient_aisha body — the shipped asset was built on a base that no longer matches the regeneration path. The sweep therefore bakes on the current base and splices into the shipped GLB (the C1 precedent); a future rebake of the adult asset on the current base would change its body identity. Not carded.
• The gown's skirt/waist (y 0.53-0.99) carries 11.7-16.6% sharp edges that render acceptably (per the The bodice fold amplitude is 34 mm on a 10-22 mm standoff, so the wave valleys land inside the body #714 grade) — the sharp-fraction metric counts rim/skirt creasing that is invisible at learner distance; clause (1)'s ceiling may be unattainable for any gown-shaped garment in this pipeline via fold changes alone (fold-zeroed estimate ~5.0%). Reported as a finding, NOT a reason to weaken the clause.
• The near-zero-area triangle count rises as k falls (2480 at k16 -> 2844 at k4): a gentler fold leaves more near-coincident subdivided triangles. Wire meshopt decimation into the humanoid pipeline — the silhouette holds at 2.81x and the station comes under budget #695 owns decimation.

CLAIM: A full k/amp sweep (k in {4,6,8,12} x amp in {0.030,0.034}, plus the k16 control, measured on the shipped asset configuration via gown-only splice) shows the sharp fraction responds monotonically to the wave count but saturates at 15.46% — NO cell clears the 3.47% ceiling; the wave-depth ratio reading is partially refuted (a k-invariant non-fold floor keeps the gown above the ceiling even with a smooth fold band, and the #714 clamp is not the floor); the sweep table and the labelled native-resolution contact sheet are committed on tracked paths.

NOT TESTED: the fold response on a robert-bodied regeneration base (the shipped gown was built on an older base than the current bake input; the ±2-point body-identity scale difference is immaterial to the 4.5x margin); decimation (#695 owns it); whether the skirt/waist construction can be smoothed without changing the garment's silhouette; the visual grade of the contact sheet — the orchestrator grades the pixels (the gown-blue fractions above are machine validation only).

[gidich]

The sweep is complete and it says no cell reaches the ceiling. It also produced a finding that
reframes this card: the fold is roughly half the problem.

The sweep table, from the worker

cell	sharp >60°	near-zero-area	peak radial
k16 / 0.034 (control)	20.80%	2480	0.4575
k12 / 0.034	18.45%	2547	0.4575
k8 / 0.034	16.67%	2747	0.4575
k6 / 0.034	16.01%	2648	0.4575
k4 / 0.034	15.56%	2844	0.4575
k4 / 0.030	15.46%	2795	0.4575

Monotone in k and saturating. The best cell sits 4.5x above the 3.47% ceiling. Peak radial excursion
is identical across every cell, so nothing here flattened the gathers.

The finding, measured by me on the SHIPPED assets

Per-decile sharp fraction over each garment's own y-range:

decile	gown	scrub top	open cardigan
d0 hem	13.0%	11.7%	10.2%
d1	10.4%	1.7%	4.4%
d2	11.1%	1.2%	3.8%
d3	9.8%	0.0%	7.4%
d4	15.0%	0.0%	2.0%
d5	29.6%	0.0%	0.0%
d6	22.5%	0.0%	0.0%
d7	20.8%	1.5%	4.4%
d8	19.5%	0.0%	0.0%
d9 collar	11.4%	0.0%	0.0%

Three things fall out, and only the first was known.

The hem is normal. All three garments run 10–13% at d0. Rim creasing is ambient and is not a defect.

The gown's SKIRT is not normal. d1–d3 sit at 9.8–11.1% where clean garments read 0.0–7.4%. The
fold wave does not run there — the fold band is d4–d8. So roughly ten points of the gown's sharpness
exist in geometry the fold never touches.

The fold band adds on top of that floor, reaching 29.6% at d5, and k moves only this part.

So the gown carries two independent contributions of similar size: a shell-wide floor around 10% and a
fold-band excess of another 10–20 points. Driving k to 4 removes part of the second and none of the
first, which is exactly the saturation the sweep measured.

MY GRADE: the contact sheet is ungradeable, and that is my brief's defect

tools/openclinxr/evidence/issue-747/bodice-band-contact-sheet.png, viewed at 2000 px from a 5010 px
source. Every one of the nine tiles is framed on a forearm and a narrow strip of the gown's side seam.
The sternum — the region this card is about — is not in frame in any tile. I cannot pick a cell from
it.

What is visible does show the k response: the k16 tile carries a dense sawtooth along the visible side
edge and the k4 tiles read smooth there.

The machine validation passed because it measured gown-blue pixel fraction per tile (18.5–20.1%),
which confirms gown pixels are present and says nothing about which gown pixels. I asked for a
"labelled contact sheet of the rendered bodice band" and gave no per-tile checklist that a wrong crop
would fail. A cell like sternum_visible: yes | no would have caught it before the render.

It does not change the decision: no cell clears the ceiling by 4.5x, so none of them ships.

What this card should become

Clause (1) asks the gown to reach the ambient ceiling. With a ~10% floor in geometry the fold never
touches, that is not reachable by fold work, so the clause as written cannot be satisfied within this
card's scope. The split is the fold-band excess against the shell floor, and I will do it rather than
ask a worker to work around it.

The sweep is preserved at salvage/issue-747-sweep (f75dd19a); the earlier C1 attempt remains at
salvage/issue-747-c1-wave (bb461d64). Neither is landed.

CLAIM: lowering the fold wave count from 16 to 4 reduces the gown's sharp-edge fraction from 20.80% to
15.46% with the gather depth unchanged, and a floor near 10% remains in the skirt where the fold wave
does not run, against 0.0–7.4% for two clean garments in the same bands.

NOT TESTED: what produces the shell floor. Whether the fold-band excess alone could reach a clean
garment's fold-free bands. And the appearance of any swept cell at the sternum, which the contact
sheet cannot show.

[gidich]

Split at ccc9fed0. This card now owns the fold-band excess only.

Clause (1) — the whole-garment ambient ceiling — moved VERBATIM to #750, because the sweep proved it
unreachable here: the gown's skirt reads 9.8–11.1% where no fold runs, against 0.0–7.4% for two clean
garments, so no fold parameter can bring the whole garment to 3.47%. An inverted guard sits in its
place and reds if #750 stops asserting it or widens its bound.

Clause (1b) is this card's new RED and it is self-referential on purpose: each gown's fold band
(d4–d8) against ITS OWN skirt (d1–d3). Measured today that is 29.6% against 10.4% on the adult gown.
The shell floor moving in either direction cannot green it, and it dies the moment the fold stops
adding sharpness the shell does not already carry.

Counterweight (2) still floors the amplitude at 0.030, and the wave count remains unpinned, so the
sweep's k = 4 result is a legal landing if a render supports it.

One probe of mine failed on its first attempt and that is worth recording: the guard's file-wide regex
matched a SIBLING clause using identical text, so it stayed green while the moved clause was weakened.
It now reads only the moved clause's own body.