The subjects

Four subject classes, one route. Every number below carries its denominator — a coverage figure quoted without one is not a number in this repo — and every subject’s values live in its own profile and fixture, never inherited unmeasured.

W3 — the character (ACCEPTED, 2026-08-04)

Record: E08-ruling-gate0.md, 35 amendments. Fixture: canon/W3-IDENTITY.md — a test fixture, not a shipping character. Profile: profiles/character.json.
Accepted at Gate 1 on the GLB at the Director’s own zoom. Measured provenance 68.8% reference / 4.2% brush / 27.0% dilation (of valid texels, eight cameras) against the earlier rejected asset’s 28.4 / 37.7 / 33.9 — reference ×2.42, diffusion invention ×0.11.
Reach: 74.1% of valid; acceptance 92.9% of reachable at eight cameras (union acceptance is a function of camera count — quote it with the count).
The architecture result that generalises: twins register; the prompt carries identity — contradicted on eight elements, the prompt won 8/8 at 7.4× separation.
⚠ Resolved 2026-08-06: the off-surface measurement ran — 2.5840%, and its composition inverts the galleon’s (paint-depleted where the galleon’s is paint-enriched), so three of five W3 headlines move the opposite way on exclusion (reach 74.10→74.30, styled/valid 68.82→69.28, dilation 26.95→26.43). The on-surface family is now the standing cross-asset family (E10-offsurface-ruling.md Ruling 7); as-recorded stays beside it everywhere. Historical: “the measurement is queued, Director-gated” — fired 2026-08-06, both tasks.

The galleon (ACCEPTED, 2026-08-05)

Record: E04-ruling.md, 29 rulings + the close (the count read “28” until E15’s verifier counted the record itself — Ruling 29 is the asset-2 blob-bound halt). Fixture: canon/GALLEON-IDENTITY.md (thirteen G-elements — G13 arrived unprompted at the mastheads and was named post-acceptance, because an element not named in the prompt will leave the same way it arrived). Profile: profiles/ship.json — 81/81 flags decided, the registry’s first clean sheet.
Accepted at Gate 1 on the five-column sheets. Mix 36.89% reference / 6.87% brush / 56.24% dilation of valid texels = 86.4% of the pre-registered 42.72% reach ceiling (a ship hides most of itself from eye level — geometry, not regression; read against the ceiling, never against the character’s raw mix). Zero credits across every generation in the arc.
⚠ Restated 2026-08-05: 2.4967% of the bake’s uv-valid texels carry positions off the mesh, and the population is disproportionately painted. On-surface the same quantities read 42.25 / 36.68 / 86.8 / 56.44 / 6.89 (E10-offsurface-ruling.md). ⚠ Resolved 2026-08-06: W3 now carries the same measurement, and the on-surface family is the standing cross-asset family (Ruling 7) — the stronger ground being that the composition inverts between subjects, so the phantom 2.5% is subject-dependent noise, not a constant bias. Historical: “as-recorded remains standing until W3 carries the same measurement” — the condition fired the next day.
Carries the waterline as data (E10): waterline_z in profile + manifest, the GLB whole, the scene’s water doing the hiding. And the native per-texel owner channel — the first asset with one.

The dragon (ACCEPTED, 2026-08-07 — designated 2026-08-05)

Record: E12-ruling.md, Rulings 1–30 + the Gate-1 acceptance; Gate 0 report. Fixture: canon/DRAGON-IDENTITY.md (eleven D-elements; ivory is the HEAD’s family since the Ruling 12e palette correction — this line read “the bone-ivory family is this scheme’s gold” until then). Profile: profiles/beast.json — 83/83 subject-data flags decided, 0 UNDECIDED (certified at handoff 15; this line read “80/81 with the one open flag deliberate” mid-arc — stale twice over in one file, caught by the E15 steward’s class sweep).
Designated: candidate 00003 (“3 is the winner, but they all look great”), with the throat crevice and the folded-membrane pinch field designated-in as named subject facts.
Gate 0’s topology headline: 9 welded shells, 99.96% of faces in the largest — an order of magnitude more connected than either prior subject (character 40–191, ship 237–512); every satellite located (fangs); membranes are closed slabs that pinch, not sheets that hole; reconstruction does not preserve the concept’s pose.
Banked so far: reach ceiling 50.46% of 3,240,510 valid texels at the eye-level eight (elevated ruled NONE — +1.77 points against a ~50-point self-occluded deficit); allocation NONE on the mesh’s own head evidence; the thin-structure constraint (the character’s thin_extent would withhold a third of the visible animal and 60% of the wings; no global value separates membranes on a subject where most detail is thin); backdrop ruled plain lavender-grey — W3’s grey scores under the key’s own cut here, bound by the membranes.
The first styled pair (0 credits) was rejected at the Director’s eye — and the rejection ruled the style register subject data (E12 Ruling 10): the beast runs ultra-realistic with no LoRA (the painterly saltroad register had arrived by inheritance), and the canny pair is being derived per-subject after the control was measured carrying a third of the clay’s own relief. Studio plan: style-registers.md.
The re-pair (0 credits) CONFIRMED the register — “a lot better” — and the per-subject canny 0.05/0.10 is ratified into the profile (Ruling 11): on grey-on-grey clay HIGH is the structure lever at ~10× low’s effect, and the ruled pair carries 108,887 control px on view 1 against the falsified 0.4/0.8’s 50,631. The pair is not yet accepted. Two named defects and the Director’s own question gate it: the missing tongue goes to geometry first (Gate 0 saw a tongue on 00001/00002 and not on 00003 — if the mesh carries none, no prompt lands one); view 5’s two material misses (pale-tan haunch on D1’s surface, bone-ivory membranes on D3’s) spend the bounded re-roll, seed the only delta; and his “could the face be more defined” reopened allocation by Ruling 2’s own clause — the ruled ladder is resolution first (a head-crop companion generation at the measured head box — never a projection reference), geometry second on his sentence only (a bust-crop re-reconstruction replaces the designated mesh: a Ruling 1 re-open, not an arm).
Handoff 5 ran all three at 0 credits (report): the mesh HAS a tongue — main-shell, route-visible; Gate 0’s omission-read falsified in one render at mouth scale — both view-5 misses resolved on the seed alone, and the resolution rung split by element: muzzle plates, nostrils and tooth rows define at bust scale; the eye is geometry-limited (brow plates, no lens recess on the clay; the denser control produced less eye than the sparse one). New banked fact: D9/D11 swap surfaces with scale — wine took the cavity at pair scale, slate took the tongue at bust scale; neither reliably owns its surface. The Director’s eye then caught the fixture’s pale-bone family — “bone” ×5 in the prompt — rendering as exposed skeleton on legs, tail underside and wing arms under the realistic register, on both seeds (Ruling 12): canon corrected in place (D2 olive-tan, D6/D7 charcoal; ivory now the head’s family), regeneration dispatched as a new decision bundle (handoff 6).
Handoff 6 ran at 0 credits and the bone reads leave at the worst seed (Ruling 13): same seed, same control, prompt only — whole-figure pale-bone 28.71% → 7.69%, tail underside 21.73% → 2.60%, the rib element gone; the seed move alone had bought only 13.00%. Banked finding, labelled: a colour term reaches structures that RESEMBLE what it names — pale ivory fangs painted view 1’s claws ivory over charcoal claws in the same string, while view 5’s stem (zero ivory words) took charcoal claws at the same seed; D5’s cheek fan took charcoal the opposite direction.
THE PAIR IS ACCEPTED (Ruling 14, 2026-08-06: “I accept. Very good!”) — view 1 v5 + view 5 v5, seed 770700, prompts E12-pair-5, 0 credits across the entire pair lineage; the three named deviations accepted at the pair with fixture dispositions landing on the bands measurement. The register row closes as earned (ultra-realistic, NO LoRA).
The bands are RULED against the accepted pair (Ruling 15): warm-olive 85.4–147.3 adopted (7 clusters, 55.23%); blue-violet suspended on one cluster — a realised D3 stratum that also caught Ruling 8a’s premise stale (the backdrop word stands on the pair’s own anchor; its realised C* 11.0 sits under the gate’s 12.0 floor); D3 carries no hue band (neutral strata, tightest landing in the table at C* 1.3); chroma floor 12.0 inherited; both dark bands are colour matches to shadow (colour-not-placement); a pooled landing table cannot clear a per-view deviation (D7’s LANDED rested on one view). H1/H4 resolved: the register was not the cause — the 41.9° collapse recurs; third subject, third time. D8 CLOSED: 193 px slit-pupil ember iris, one blob, in the head box, on an accepted artifact; the bust-companion contradiction stays parked with the allocation ladder. Dispositions: D5/D7 canon unchanged, deviations recorded.
The gate’s pair-validation halted itself before gating anything (Ruling 16): the pre-registered third branch fired on S-occlusion realised — the fixture’s seam stressor as cool marginal-chroma crevice shadow (7,293 px, 121 fragments, largest 1,342; hue ~234 at C* 14.4, 87° from any band). Ruled: floor stays 12.0 (the floor-14 trade deletes 6.6% of in-band mass), no band widens, no element is added for a seam; the per-pixel gate and the cluster table are different instruments (first measured disagreement — a 3,927 px cluster floor cannot see a diffuse 121-component population). The gate runs report-structured with null bounds — the galleon’s own published configuration — and the halt trigger is the E07-class signature (one coherent garment-scale component) at the advisor’s eye.
The eight twins ran at 0 credits (Ruling 17): twins 1/5 pixel-identical to the accepted pair (0 of 1,835,008 differing — the recipe reproduces its accepted artifact end to end); the resemblance channel resolved by natural experiment (the bleed rides the fangs term — foot ratios 0.39–0.77 with it, 0.02–0.11 without); the hue gate’s cleanest twin hid a 43,999 px flat-black foreleg (achromatic = invisible to a chroma-floored gate by construction — the achromatic channel is now permanent), resolved by the one bounded re-roll (seed defect; the calibration opposite of the canon-caused D7 case). The Director’s eye then caught the wing skeleton bone-ivory on the two wing-spread views: D1’s surface by fixture assignment, never named in the prompt — named now, and the naming saga that followed (compound term half-binding, the split, the measured discovery that term binding is seed-dependent) is Rulings 17–20.
The exemplar arc (Rulings 20–23): the Director raised the bar — these sprites inform the datasets and will be displayed — so the set regenerated whole from the occupancy-complete canon (his nape catch triggered the full audit: every modelled structure → one owner, table in the fixture). E13 Gate 0 passed at 0 differing pixels (crop-camera projection capability proven). The membrane word had its one iteration and falsified itself against its own pre-registered band (opacity improved, warmth stayed — a term cannot deliver opaque-and-cool); warm membranes accepted. His “not very consistent” was measured exact: the two views he named are the set’s tonal extremes, 7.8 L* apart — the harmonization pass (identity-tested Lab transfer toward reference view 1) is adopted as projection’s input. Stage 1 + the E13 head-crop pass dispatched (handoff 13) with a pre-registered A0/A1 comparison → sweep gate → strokes → Gate 1 at the exemplar bar.
Stage 1 is BANKED and the crop pass ruled (Rulings 24–25): A0 = 1,430,687 styled texels = 44.2% of valid = 87.5% of the pre-registered 50.46% ceiling (between the ship’s 86.4% and the character’s 92.8%), born on-surface. The E13 crop pass was not adopted — crop-owned regions read off-register on the judging sheets (the third measured instance of frame-changes-register: companion orange → crop gloss → crop scarlet) — while its capability is banked permanently (crop-camera projection at zero changed pixels; the frame-bounds and width-fit fixes; reference-conditioning is the named future lever). A2’s finding: the 4096 atlas under-resolves the head at BOTH frames (11.07× against the crop paint); the real lever is a per-subject head rect in the bake — specced, waiting on the Director’s timing, and it re-anchors the ceiling if armed. The stroke lane: thin_extent 0.005 global on the artifact criterion (the arc’s last UNDECIDED closes), elevated cameras closed against their pre-registered falsifier, four strokes in spiral order closing 71.60% of 204,617 reachable holes. The registry sweep then caught the advisor clearing a block in prose (Ruling 26) — a ruling that decides values pays for them in registry entries — both transcriptions landed, the sweep re-ran to 0 UNDECIDED (83/83 decided), and the strokes are released. Handoff 15 runs: strokes → finalize → pack → the five-column sheet → Gate 1 at the exemplar bar.
ACCEPTED at Gate 1 (2026-08-07, “Fantastic!! This is really good!” — Ruling 28, on the GLB at the Director’s own zoom, including a top-down rear whose surface only the dilation fill could have painted): four no-LoRA strokes (99,643 texels, zero re-rolls, zero credits across the entire arc), the fifth brush signature measured and named forward (dark desaturated crevice fill), mix 44.15 / 3.07 / 52.78 of valid = 87.49 / 2.86 / 9.65 of reachable. The three items named at the gate closed accepted-at-the-asset; the head-rect allocation arm stays parked at his timing. All four then ran: the dense export (handoff 16 — dataset asset #3, the first manifest under the lane’s 1.3.0), the ingest, the E15 context index, and the E14 arc.

The longsword (ACCEPTED, 2026-08-08 — designated 2026-08-07)

GATE 1: ACCEPTED (“Fantastic! I accept”, ruled on longsword_hero.glb at the Director’s own zoom — Ruling 32). The route’s fourth accepted asset and fourth subject class (character, ship, beast, prop), at zero credits across the entire arc. Final mix 45.25% reference / 2.07% brush / 52.68% dilation of valid — within 1.1 points of the dragon on every class — and 88.71% of the rendered visible surface is the accepted pair’s own paint. The stone that drifted was returned to the identity by arithmetic (the garnet re-projection, Rulings 25–26 — a hue rotation of the four drifted twins’ stone regions, L preserved, zero generation), the one mask-boundary defect repaired on an asserted-count union (Ruling 28), all eight strokes committed with both edge-ons clean on the 20b misbind watch (Ruling 29), finalize’s fill replayed byte-for-byte (Ruling 31). Every E14 artifact is citable-only.
Record: E14-ruling.md, Rulings 1–35; Gate 0 report; handoff-2 report + Task-1 sweep; handoff-9 report (finalize/pack/Gate-1 staging). Fixture: canon/LONGSWORD-IDENTITY.md (five L-elements — the compact case; occupancy audit at authoring, day one). Profile: profiles/prop.json — swept at handoff 2 (3 UNDECIDED, the pre-stated three); the canny pair (0.10/0.25, Ruling 6) and the backdrop (plain lavender background, Ruling 7) ruled per subject; thin-extent decided 0.0 at the stroke-lane ruling (Ruling 24d); finalize SPENT at Ruling 31g.
Designated: candidate 00001 (“00001 is my favorite”) — 1 welded shell at 1.000000, zero boundary edges (the cleanest topology the route has recorded, on all three candidates), 121 non-manifold edges with the grip wrap as the pinch locus. The softer pommel apex and lumpier wrap are designated-in as subject facts.
Register ruled day one, chosen fresh (Director: “Ultra-realistic, no LoRA”) — the style-registers law executed at designation for the first time; the fixture was authored under the register (every colour word names a material that IS its colour), applying the dragon’s 12e lesson at birth instead of after a rejection.
Gate 0’s class headlines: the route’s first portrait subject (widest-horizontal/height 0.2258 — the framing family pins HEIGHT here, per subject, where the beast pinned width); THE HOLLOW FINDING — every reconstruction the route has made is a double-walled shell, walls ~two voxels, measured with the accepted dragon and galleon as controls (route-wide, E14 Ruling 3, CLAUDE.md constraint extended); and the pinch prior — fine relief becomes non-manifold pinching at constant density, not denser mesh, so allocation ruled NONE on the lowest density contrast the route has measured (1.135×).
Measured at handoff 2 (the pass before any generation): pre-registered reach ceiling 51.33% of valid at eight cameras — a topology fact, not a camera-count fact (93.34% of unreachable texels are the hollow finding’s inner wall; 24 cameras buy 51.86%; the instrument’s ray bias overstates by a recorded +0.97 points, Ruling 10b); off-surface at birth 11.0875% — the fourth point that broke the cross-subject class reading (E14 Ruling 9: the population is the bake-margin ring on all three measured bakes and the rate tracks island count — this subject packs 46,496); elevated cameras ruled NONE at converged ray density (+0.12 points for the best pair; the convergence itself sharpened the 7b law — the profile-frame grid understated up-facing reach 3.9×, Ruling 10a); the thin curve banked with its face-on/edge-on inversion (Ruling 10d); the ship’s missing fourth-consumer framing pin found by the cross-profile sweep and completed (Ruling 11a).
The styled target pair is ACCEPTED (Director, 2026-08-07: “I agree with view 0 being the best. I love it. You have my acceptance.” — E14 Ruling 16). Task 4 ran at 0 credits: view 0 at seed 770700 (the pair’s primary), view 1 at 770701 after the first roll was rejected against the fixture’s occupancy audit (gold across L2’s crossguard) and re-rolled once, bounded, with the rejected artifact preserved — the specification working (Ruling 12). All five elements landed on both accepted views. Along the way: the collar rings earned their term (gold collar rings, Ruling 13); the ruled word materialised in the magenta band doing its job by 8–11× over the key’s cut with steel arriving dark — word stands, grounds corrected in place (Ruling 14); the one-string premise failed narrowly — per-view stems, boss term off views 2/6 (Ruling 15). The recipe anchor is converted; the estimates file is superseded by realised values.
Named and parked at his timing: THE ACTIVATED STATE (Ruling 16c) — a second authored state (glowing gem, forge-hot steel, no flames) as the E10 waterline layer mechanism with the contact query swapped for the occupancy map. Its time is now named (Ruling 35, 2026-08-08): it opens INSIDE the polish arc, after the four MCP tools pass their tests — the fixture authors from scratch when opened, per-state derivations re-run.
The bands are RULED (Ruling 17): floor 12.0 — an inherited number vindicated at the measured knee (the chroma density is monotone; no antimode exists, and the executor refused to dress one up); wine 332–32 merged (forced by continuous support — the 60° span is the gem’s instability made numeric) + gold 42–104; the lavender-rim band admitted as a rim-admission, not a material (placement-proven: 93.1% of it within 2 px of the boundary), with a standing depth diagnostic; the gate report-only — its only known-bad artifact fails by occupancy, which colour-not-placement cannot see, so it has no validation target on this subject (demonstrated, banked).
The twin set measured its seed (handoff 3 Task 2 → Ruling 18): at 770700 all four diagonals sprawled gold across L2’s crossguard (93–96%, 0.0% on both face-ons — categorical, and the pair’s own rejected-then-accepted view 1 had already convicted the seed and cleared 770701); both edge-on views hallucinated a face-on guard from a ~5,600-px control — view 6’s re-roll clean (IoU 0.93), view 2 a recorded second failure, excluded from stage 1 (cost measured: 2.8 points; the brush stage owns the hole). The depth diagnostic’s first run caught both corrupted views independently; zero credits across ten submissions.
The re-roll pass resolved to the clean exit (handoff 4 → Ruling 19): all four diagonals landed iron at 770701 — exactly zero gold pixels against 93–96% at 770700, the zero validated against the replaced artifacts on the same code path. The seed’s price: the L5 gem drifts magenta at 770701 (six 770700 artifacts garnet, five 770701 drifted, no exceptions — the accepted pair’s own view 1 included), so the seven-twin set is L5-non-uniform between view groups; stage 1 carries a gem-region readout and the brush stage owns the repair if the blend fails the eye. Also banked: registration failure on this subject is a bbox phenomenon, not an IoU phenomenon (the populations overlap; a rejected roll out-scored an accepted twin), and the depth diagnostic’s blindness has two faces (drifted material, not just backdrop bleed — second catch in two runs).
The Director’s eye overturned an acceptance (Ruling 20, 2026-08-08): view 6’s re-roll carries a face-bearing mass at the crossing — the excluded view 2’s death’s-head has a quieter sibling, and the advisor had accepted it on registration numbers and a downscaled strip (the contact-sheet error, owned in the ledger with two more misses). Both edge-on views are now excluded; the edge-on finding hardened to zero identity-clean rolls in four — ornament amplitude scales inversely with constraint density, and at edge-on constraint it goes figurative. The edge-on surfaces belong to the brush stage by construction. Registration measures WHERE paint sits, not WHAT it is.
Stage 1 ran as SEVEN twins (the session launched before Ruling 20’s amendment; the Director let it complete as the measured comparison — Ruling 21): styled 47.22% of valid = 92.09% of the pre-registered 51.27% seven-camera ceiling (the beast banked 87.5% of its own). Its findings: a marginal is a property of an ordering, not of a camera — Ruling 18c’s 2.8-point price for view 2 was a ladder-position artifact, true set-level cost 0.0634 points (43.8× overstated; 20c’s v6 price corrected the same way); the projected gem is a garnet core inside a magenta shell split exactly on the ownership partition (one violet stone to the eye — the Director’s window holds the L5 verdict); the crossing is the worst-covered structure (38.1%); the background probe’s corner-median reference is the retired keying method’s last live consumer (errand batch); the brush’s territory is 148,470 reachable-unstyled texels = 4.05 points.
STAGE 1B IS THE BANKED A0 (Ruling 22): six twins, styled 45.25% of valid = 88.71% of the pre-registered six-camera ceiling (beast: 87.5% of its own) — the contamination gone at the walk (the crossing’s smear absent in the after column), repartition exact, zero credits across the whole arc. The re-projection’s finding became the price law’s second clause: a reach price is not a paint price — view 6’s exclusion cost 0.27 points of reach but 1.97 of paint (7.4×; 94% of the loss is reachable-but-unpainted — erosion and trust bind where a subject is thin), and the two edge-on cameras failed by different mechanisms (angular vs occlusion). The brush’s territory is the blade’s edge-on centreline ribbon + the guard seam: 5.76 points.
The gem is GARNET (Director, 2026-08-08, one word — Ruling 23): the fixture holds; the stone’s drifted colour joins the brush as a named stroke class (the garnet repaint, ~70% of the stone’s above-floor colour, the garnet core kept). The 19b term question (“deep red garnet”?) goes to the stroke-prompt derivation with the 12e law in the room.
The lane ran to acceptance and the arc CLOSED at Ruling 35. The stroke lane was ruled at HALT 1 (Ruling 24: eight strokes merged with the garnet repaint, spiral order, thin_extent decided 0.0); stroke 1’s double misbind took the stone off the generation path (the garnet re-projection, Rulings 25–27, zero generation) with the collar repair on an asserted-count union (Ruling 28); HALT 2 was accepted (“I love it”, Ruling 30) and Gate 1 followed (Ruling 32 — the first bullet above). Then the same day: the dense export at zero gap notices (26 cameras, the six-class provenance palette — a route first; Ruling 33); the ingest — dataset asset #4, live digits = dry digits line for line, the lane pushed to main at the Director’s word, the dataset now four dense assets / 114 records across five ingests (Ruling 34); and the polish arc CHARTERED (Ruling 35): every accepted exemplar polished after the four MCP tools are built and test-verified at the shipcheck bar — dogfooding the tools, re-informing the dataset, the humanoid re-made photo-real without the saltroad LoRA, and the activated state opening there.

The calibration table this page exists for

quantity	character	galleon	dragon	longsword
welded shells	40–191	237–512	9–12	1–331 (00002’s 331 = inner-wall fragments, not detail)
widest-horizontal / height	~0.46–0.72	1.04–1.11	1.61–1.74	0.226–0.358 (first portrait)
reach ceiling (of valid, as recorded)	74.1%	42.72%	50.46%	51.33% (topology-bound: 93.34% of unreachable is inner wall)
off-surface at birth (>1 px) ⚠	2.5840%	2.4967% (full bake)	2.6430%	11.0875% (46,496 islands)
off-surface composition	paint-depleted (stage-1 1.94%, reach 0.899×)	paint-enriched (stage-1 3.06%, reach 1.426×)	no consumers yet	—
accepted mix, as recorded (ref/brush/dilation)	68.8 / 4.2 / 27.0	36.89 / 6.87 / 56.24	44.15 / 3.07 / 52.78	—
on-surface restatement — the standing family	74.30 reach / 69.28 styled / 26.43 dilation	42.25 / 36.68 / 86.8 / 56.44 / 6.89	born on-surface; 87.49 / 2.86 / 9.65 of reachable	—

⚠ The off-surface row was re-read 2026-08-07 (E14 Ruling 9): the population is the bake-margin ring on every measured subject and the rate tracks island count — the three-subject 0.15-point agreement was island-size coincidence, so this row is quoted per subject with its island count, never compared bare across subjects. The on-surface family’s arithmetic is unchanged.

Three lessons live in this table. Subject properties do not interpolate — the dragon’s shells and reach landed outside any band the two priors suggest, in opposite directions, and the arc’s standing rule is that a new subject class has no working prior. Bake artifact classes transfer at the RATE and not the COMPOSITION — the off-surface rate replicates within 0.15 points across three subjects, while its class composition inverts between the two measured ones; carry the rate, never the split. And the population is bounded: 90.85% of it is the mask’s single outermost texel ring, largest blob 33 texels — measured, localized, no route change warranted (E10-offsurface-ruling.md Ruling 7c).