decisions/S6_R7_AND_FWHM_RULINGS_2026-08-10.md
**2026-08-10, director (milks sixteenth seat, claim 1ffa183d). The two rulings the
second-succession owed board minted (docs/PIPELINE_LEDGER.md:6060) and the S6 stage-proving run
named rather than smoothed (build/3d/best_a14/retopo_s6/S6_WRAP_STAGE_PROVING.md, THE TENSION
block and R-3's FINDING FOR THE DIRECTOR). Both run the standing decision protocol: steelmanned
alternatives, deep-dive against the landed numbers, adversarial pass both directions per the
care-calibration law, recommendation APPLIED. Binding on runs adjudicated AFTER this lands — no
live lane's pre-registered terms change. No threshold is loosened anywhere in this document: the
C4 floor 0.0314159265 stands, the eight-decimals identity bar stands for the class that can meet
it, and the ≥ 3-across-FWHM number survives both rulings intact.**
---
R-6: *"Auto-retopo is the right tool for props and creatures; it is the wrong tool for a hero
character's deforming regions"* (CRAFT_AI_3D_GENERATION.md:527-528, the practice law restated
:686-696; charter §S6). R-7 as written carries TWO halves (CRAFT_AI_3D_GENERATION.md:700-705):
the PRECEDENT — *"the A2_CLEAN precedent — identical to A1_GENERATE to eight decimals on all four
channels and all three views — is the house standard"* — and the RULE — *"Any topology-changing
stage publishes the same comparison through p1_score.py against the C4 sensitivity floor."* The
law never says what "the same comparison" MEANS for a stage that replaces every vertex: identical
to WHAT baseline, delta read HOW. The stage-proving run's own words: **"a wrap cannot be identical
to eight decimals; it replaces every vertex. R-6 (wrapped template) and R-7 (prove it is free) are
in tension as written"** (S6_WRAP_STAGE_PROVING.md, THE TENSION; instruments.py I4 recorded it
before the run). This ruling defines the comparison rather than rewriting the law.
1. Status quo — the eight-decimals precedent read as the bar for every stage. The strongest
reading of the original law's precedent half: any stage that cannot prove eight-decimals
identity is not trusted. Consequence, measured: the S6 wrap can NEVER pass — not a hard wrap,
not a perfect one — because vertex replacement is the stage's definition. This bans the stage
the S6 RULING adopted on forcing evidence (charter :1214-1226, OPEN-15). The two director
rulings would then contradict, and the practical result is the worst state on the board: the
stage runs anyway (it is adopted) with NO free-proof at all governing it, because the rule
half's "same comparison" stays undefined for a stage with no identity baseline.
2. A higher floor for wraps. Loosen the C4 floor for topology-replacing stages since they
"can't help moving things." REFUSED on its face — the loosening direction is the banned
direction; a stage that moves the silhouette more than the C4 floor has moved likeness, whoever
moved it.
3. Exempt S6 from R-7 and rely on conformance p95 alone. The stage-proving run already
falsified this: conformance p95 read 0.2867 cm while the back-view silhouette was +15.65× over
the C4 floor — *"a conformance p95 of 0.2867 cm and a silhouette delta of 15.65× the floor are
not in [agreement]. That is precisely why R-7 exists as a separate proof"*
(S6_WRAP_STAGE_PROVING.md). Conformance measures distance-to-surface; the lens measures what
the plates see. They disagree exactly when it matters.
4. Two-class restatement (RECOMMENDED, APPLIED). R-7 splits by what the stage claims:
dilation): the bar as written. Identical to eight decimals, clean_shells.py precedent,
unchanged in every particular.
vertices): the paired-baseline method the stage-proving run used, and the only resolution
the evidence allows (its own words): score the wrap AND the subject it wrapped against the
SAME plates under the SAME lens in the SAME run, and read the per-view DELTA against the C4
floor. Bar: |delta| < 0.0314159265 on EVERY view under EVERY convention (AS_SCORED,
C1_CLIP_GUARD_DISARMED, C2_ISOTROPIC_NO_ALIGN — the ROUTE_A_RESIDUALS.md FINDING 4
roster; all three published, none picked). Views are never averaged and an improvement on
one view never buys another — the S5 side-likeness lesson (ANATOMY_REFIT_V2.md:202-205).
Baseline = the subject scored fresh in the same run, so the subject's number is pinned and
"both moved together" cannot hide.
silhouette holds? R-7b governs the FREE-proof only — it is NECESSARY, never sufficient. It does
not absorb or replace the stage's other instruments: quad fraction, loop presence, I3 landmark
edge density, the face-quality row R-4 minted, topology_bar, conformance p95, and the
wireframe judge read all stand independently (*"a number can be right and the render still
wrong… Both judges, or neither"* — charter :104-107). A wrap that passes R-7b and fails any of
those still refuses.
Keeping R-7a for wraps is not the "strict" option — it is the NO-BAR option, because the
adopted stage would run ungoverned while a dead bar points at it. The two-class restatement is
the strict option: it puts the same C4 floor every repair and rig rung answers to onto a stage
that previously had no floor it could even fail against.
The C4 floor was minted for view-level silhouette deltas of position-only edits; applying it
per-view to a whole-topology replacement is a *reuse*, not a derivation, and a wrap could in
principle sit just under the floor on all views while re-sampling badly everywhere. Answer: that
failure mode is exactly what I3/R-4/topology_bar own (sampling and face quality are their rows,
not the lens's), and the floor's reuse is the same one-ruler discipline the rig row already
practices — one floor, one lens, every stage measured with it. A per-stage bespoke floor is the
gate-fitting direction.
**R-7 is restated as a two-class law. R-7a (position-preserving): identical to eight decimals,
unchanged. R-7b (topology-replacing): paired-baseline per-view delta under the C4 floor
0.0314159265, all three conventions, no averaging, baseline scored in-run; necessary never
sufficient — the topology instruments stand independently.** The stage-proving run's resolution
is hereby the law it improvised. Applies to runs adjudicated after this lands; the live
production-wrap lane's pre-registered terms are unchanged (they are this method already).
---
topology_bar I3 requires ≥ 3 vertices across the clavicle-ridge FWHM (0.733 cm; OPEN-15,
charter :1219-1221). The wrap read 0.3179 against the subject's 1.0762 — *worse than the subject
on the exact number the S6 RULING turns on* — and the run's sampling ladder proved the bar
unreachable by global density at any hero budget: Catmull-Clark level 2 is already 214,048
quads and reads 1.3506; even 856,192 quads extrapolates to ~2.70 < 3
(S6_WRAP_STAGE_PROVING.md R-3, C_THE_SAMPLING_LADDER). The run's FINDING FOR THE DIRECTOR
poses the fork: either the ridge is a displacement/normal-layer feature and the bar belongs to
S7's displacement source, or the bar as written cannot be met by S6.
1. Keep ≥ 3 as a global base-cage bar. Honest consequence: a permanently RED row no hero mesh
can clear — 856k quads is not a hero base (the run's own finding). A bar that cannot be met is
not a bar; it is a stage ban wearing an instrument's clothes, and it contradicts the adopted
S6 stage the same way alternative 1 of Ruling 1 does.
2. Move ≥ 3 wholesale to S7's displacement source. Trivially met — at the R-11 texel bands
(512/1024/2048 px/m by class, CRAFT_TEXTURE_LOOKDEV_GROOM.md:274-278), 1024 px/m puts ~7.5
texels across 0.733 cm — but it silently decides that the clavicle ridge is NOT geometry, and
the S6 RULING's forcing evidence says the opposite: *"no displacement pass can carry
micro-anatomy on this topology"* (charter :1220), and the MUST-READ register requires the
clavicle *"full length with S-curve and both ends findable"* as a soft ridge under a 5.3 mm
mantle (CRAFT_ANATOMY_STANDARDS.md:325-327). A ridge that exists only in a normal map
vanishes at the silhouette and at grazing light — the exact Josh complaint ("no Adam's apple
or collar bone", ROUTE_A_RESIDUALS.md:26) returns.
3. Local refinement at named landmark sites (RECOMMENDED, APPLIED). The S6 RULING already
contains the answer in its own words: *"wrapped template topology with MetaHuman-class
edge-loop density AT deformation zones AND LANDMARK SITES"* (charter :1217-1218). MetaHuman
density is not global — it is loop PLACEMENT. The ladder measured the wrong operation:
global Catmull-Clark subdivision, which the run itself labels honestly (*"subdividing a
wrapped cage INTERPOLATES. It does not go back and look at the subject"*). A template with
loops routed along the MUST-READ ridge lines meets ≥ 3 across a 0.733 cm band for a few
hundred quads per site, not 856k globally — that is what an authored hero template IS. The
derivation of the number stands: ≥ 3 vertices across the FWHM is the minimum that SAMPLES a
bump rather than interpolating past it (two edges across the feature — the Nyquist floor for
representing one ridge in geometry), which is why the bar was set there and why it is kept.
1. **The I3 bar is re-scoped LOCAL, not loosened: ≥ 3 vertices across the FWHM, measured ALONG
THE FULL RIDGE LINE at each MUST-READ landmark site** (clavicle full length with both ends,
acromion, scapular spine/medial border, suprasternale region, iliac crest + ASIS plane
change, patella, anterior tibial crest, malleoli, olecranon/ulnar styloid — the
CRAFT_ANATOMY_STANDARDS.md:325-335 register), to be met by TEMPLATE LOOP PLACEMENT and
local refinement, never by global subdivision. Sampling along the full line kills the gaming
path (three dense loops planted at the tape point while the ridge elsewhere stays coarse).
The re-wrap at a landmark-refined cage is the geometry claim — the run's own OWED item, now
with its acceptance defined.
2. The feature class splits by SCALE, mirroring the adjudication S7 already carries
(charter :331-340): **silhouette-scale form (≥ 1 mm relief, findable in an outline or at
grazing light — the ridge's plane change and S-curve line) is BASE-CAGE GEOMETRY, S6's row.
Sub-millimetre relief (the ridge's micro-modelling, vein outlines, knuckle/pore/cavity
layer) is S7 DISPLACEMENT-SOURCE content**, where the sampling requirement transfers to
texels: ≥ 3 texels across the feature's FWHM — met by the R-11 bands and measured by
texel_density.py per region, so the row is armable the day S7 exists.
3. OPEN-15 stands discharged into these two rows. Its forcing evidence ("no displacement
pass can carry micro-anatomy on this topology") remains true exactly at silhouette scale and
is the reason part 2 routes silhouette form to the cage; the 1.08-vertices-across finding was
the symptom of a cage with no landmark loops, which part 1 now names as the defect the next
template/re-wrap must clear.
No: I3 was always a landmark-site row — the global rows (quad fraction, loop presence at
elbow/knee/shoulder/eye/mouth, hero loop budgets 6-8/3-5) are separate instruments and stand.
Does the texel transfer smuggle geometry into texture? No: only the sub-millimetre layer
routes to S7, and S7's own pre-registered refusal condition (leg E relief + the S5 age read
scored BEFORE and AFTER, any move toward ADULT refused — charter :344-346) polices amplitude
there.
reads strict while guaranteeing the stage ships ungoverned or not at all. The care law names
this direction the costlier error.
The ~few-hundred-quads-per-site claim is asserted from MetaHuman practice, not measured on this
template — a landmark-refined HM08 derivative could cost more. Answer: the ruling sets the BAR,
not the budget; if the refined template blows the hero budget, that is a refusal the re-wrap run
publishes with its numbers, and the fork returns to the director with data instead of
extrapolation. The bar's meetability is now falsifiable, which the global version never was.
---
topology_bar I3: global read → per-landmark-site read along the ridge line, same ≥ 3 number(implementation lands with the topology_bar site-arming owed row, which rides the wrap surface
a live lane holds — NOT edited by this seat).
texel_density.pywhen S7 exists).
instruments.py I4's recorded tension: resolved by Ruling 1; the file itself is a live lane'ssurface and is not edited by this seat — the ruling is the record, the wiring is the owning
lane's.