AAAAA_IMPROVEMENT_REGISTER_2026-08-06.md

pipelines/AAAAA_IMPROVEMENT_REGISTER_2026-08-06.md

THE AAAAA IMPROVEMENT REGISTER — 2026-08-06

What this is. Josh asked one question: *"what are you using that we need to improve on?"* Six

domain lanes read their own records cold — character 3D, creatures, music, animation, world and

rendering, engine tooling, garments and texture — and returned the tools actually in use, the

measured ceiling each one hits, and what would move it. This file is the synthesis: every improvement

across all six domains, ranked by slice impact ÷ cost, each row carrying the measured number and

the record it came from, its cost class, and the queue wave it belongs to.

Authority. PROPOSAL-TIER. This register ranks work; it rules nothing. Canon, the CVD, and Josh's

rulings bind every row in it. Where a row proposes a NEW wave, the wave is named so it can be

appended to docs/HANDOFF_2026-08-05_ACCOUNT_SWITCH.md section 8.7 (the standing queue) rather than

run beside it.

---

1. Executive summary — the ten-line answer

1. Four of the things we use are at their measured ceiling; most of the stack is not the problem, and the records say which is which rather than leaving it to taste.

2. TripoSG mesh lift and the ratified MV-Adapter texture route are PROVEN and are NOT the ceiling — 0.9883 texel observation on the serpent rebuild, 151 of 187 creature meshes staged, 61 of 61 lookdev values landed, GPU p99 13.282 ms inside a 16.67 ms budget with headroom and nothing to spend it on.

3. Ceiling one is RETOPOLOGY: AutoRemesher, the only quad-retopo tool in the stack, produced literally zero quads on the face (remesh_used=false, out_quads=0, CHAR_0001_FACE_finish.json), so no stage anywhere emits a shared-topology cage — and continuous aging (a RULING), every facial rig, morph-target customization and 187 aging meshes all sit behind that one missing stage.

4. Ceiling two is ARCHETYPE COVERAGE: one motion solver (LIMBLESS) serves 13 of 187 creature rows (6.95%), the elevation and marks tools are keyed to that single rig by construction, and no coat or groom stage exists at all for the 62 QUADRUPED and 11 ARBOREAL_PRIMATE rows.

5. Ceiling three is ACE-Step's PROMPT WINDOW: 56-59% of every derived music instruction is shed before generation to fit 189 usable caption tokens, the six dimensions pinned at full error are dimension-for-dimension the shed ones, and the one experiment that would separate "the generator cannot" from "the instruction never arrived" ran on the wrong checkpoint and self-declares admissible=false.

6. Ceiling four is UE 5.8's PYTHON SURFACE: it cannot create-and-register a component, wire a graph node, write a Niagara module input, or read per-LOD triangles — and we already own the fix pattern (a headless-callable UFUNCTION in Source/HumanityEditor), have instantiated it twice, and have written 17 probe scripts to rediscover the same absences instead.

7. Two instruments we currently trust are actively misleading us: the foot-slide mean (four waves proved no content clears it — Epic's own motion-matched library made the whole-run number WORSE, 0.5650 to 0.6726) and a far-field artifact probe pinned to a camera station that silently moved 3.65 km horizontally and 802 m down under an unchanged site id.

8. The cheapest wins on this register are not new tools at all — they are passes we already own and have simply not run: the protagonist's body is a flat cream CONSTANT sampling zero textures across the lower two-thirds of every frame, twelve authored creature animations are unbindable behind a patch already written out verbatim, and the pivot head's de-light pass exists and was never run on it.

9. Exactly two rows on this register need money — canopy substrate and a real hit-reaction set — and neither of them is the $1,000 music corpus, which has no lever on this generator until the 189-token question is settled.

10. Ranked by slice impact ÷ cost the top twelve rows are all free/config or a single tooling-day, and the binding constraint (the wrap cage) ranks thirtieth by ratio and FIRST by absolute impact — which is exactly why it gets its own wave instead of competing with the finish passes for the same hours.

---

2. How this register is scored

Slice impact (I), 1-5, measured against THE JOSH GATE — the AAAAA-polished PROLOGUE→CH 13 slice,

not the whole arc.

• 5 — a player looking at the slice sees the difference, or a RULED requirement is unblocked.
• 4 — a whole class or wave of the slice becomes buildable.
• 3 — a measurement that decides a route or a spend becomes trustworthy.
• 2 — a known defect closes on one asset or one instrument.
• 1 — bookkeeping and discoverability, which the anti-orphan doctrine still requires.

Cost (C), 1-5, in the four declared classes.

• free/config (C=1) — no new code. A config entry, a re-run of an existing proven pass, a marker

placement, a record emit, or a commit of a patch that is already written out verbatim.

• tooling-day (C=2, ×3 → C=3, ×5 → C=4) — new script or C++ authoring in one lane, sized in days.
• model-swap (C=3) — a changed or added generative stage: a new checkpoint, a new model in the

chain, a first install, or GPU-hours as the dominant cost.

• spend (C=4) — money, seats, marketplace acquisition, or a human-gated purchase.

Rank = I ÷ C, ties broken by absolute impact. The ratio deliberately floats cheap finishes above

the binding constraint. That is the correct read, not a defect in the scoring: the finish passes and

the wrap-cage engineering are different lanes that do not compete for the same hands, and section 5

names the binding constraints separately so they are never lost inside a ranked list.

Wave column. Q1-Q10 are the standing queue in HANDOFF section 8.7. Six NEW waves are proposed

here and named in section 6; they are beyond the floor of what Josh asked for and are NAMED as such

so they can be vetoed.

---

3. THE RANKED REGISTER

3.1 Tier A — ratio 3.0 and above (free/config, real slice movement)

3.2 Tier B — ratio 2.0 to 2.5 (one tooling-day, or free/config at lower impact)

3.3 Tier C — ratio 1.25 to 1.67 (the structural work, including both binding constraints)

3.4 Tier D — ratio 1.0 to 1.5 (real, sequenced behind the tiers above)

3.5 Tier E — ratio below 1.0, and the bookkeeping the anti-orphan doctrine still requires

---

4. What is NOT on this register, and why

Ruling work OUT on evidence is as much of a finding as ranking work in. These were considered and

rejected, each on a measurement rather than a preference.

• Buying better locomotion clips to fix the foot-slide band. Epic's motion-matched library, wired

in as ONE controlled variable, made the whole-run number WORSE (0.5650 → 0.6726) and jog and sprint

with it. Two independently authored AAA sets both floor at 0.32-0.36 at walk on their own authored

paces. The band is a statistic problem (row 12), not a content problem.

• Stride-warping automation as a way to reach the band. Arm C's cycle-unity control already

answered arm W's question: at foot_cycle_ratio 1.0050 there is no stride deficit left to remove

and 0 of 4 bands are green. Row 55 buys future AnimGraph automation, not this band, and is ranked

accordingly.

• Loosening the contact-selector window. 51 of 120 swept windows do reach the band — by collapsing

contact fraction to 0.000-0.035 against an armed floor of 0.10. A gate that passes by measuring

nothing is the failure mode this repo has already minted gates against.

• The $1,000 music exemplar spend, for now. Buying cards sharpens targets that cannot fit through

a 189-token window. The spend has no lever on THIS generator until rows 7 and 43 return, and rows 9

and 28 buy the same class of information for $0.

• A better generator as the answer to the music ceiling. Not ruled out — ruled UNDECIDABLE on

current evidence, which is the point of rows 7 and 43. The turbo config demonstrably lands the meta

channel exactly (global_bpm error 0.000 on two of three samples, total_length_s under 0.004 on

all three) and steers several surviving clauses; nothing on disk shows what it does with the other

16 dimensions because they were never in the prompt.

• Another remesher for the face. AutoRemesher is not underperforming, it is returning zero, and

the routing record already names the shape of the fix: a WRAP solver conforming a cage that already

has correspondence, not a second attempt at deriving one.

• Renderer or lighting work on the world. The CINE_TIER A/B proved the profile reaches the

renderer and the frame still reads thin; the light rig is physical with 61 of 61 values landed; the

fill A/B came back NOISE on all five metrics and the fill stays off as a ruling. The world's

ceiling is substrate (row 35) and the game thread (row 45).

---

5. The binding constraints, held out of the ratio ranking on purpose

Four constraints bind, and none of them wins on ratio. They are named here so a ranked list cannot

bury them.

• No correspondence-bearing face (row 34). The two head routes each hold exactly what the other

lacks and nothing joins them: the AUTHORED head has correspondence and parametric control

(skull_head_units 6.5673 against the canon target 6.567272) and is judged BELOW on appeal 4 of 4

plus once more on fixed frames; the PIVOT head has appeal and no correspondence at all. Everything

downstream — aging, FACS, LODs, customization, roster scale — is downstream of that one gap.

• One-body-plan creature pipeline (rows 30, 16, 47). Rig, motion, elevation, marks and the arc

field the marks register through are all keyed to the LIMBLESS chain. 13 of 187 rows served; 42

more blocked upstream on a canon ruling (row 5) that no tool can unblock.

• Conditioning bandwidth in music (rows 32, 43, 7). 56-59% of derived clauses shed, and the

dimensions pinned at full error are the shed ones. The experiment that separates the two

explanations is complete and inadmissible.

• No editor-module extension habit (rows 21, 37, 38, 48, 49, 54, 55). The project owns the fix

pattern and has instantiated it twice. Every lane instead writes a probe and ships a weaker

substitute measurement, then names it honestly — and the honest names accumulate as permanent

UNMEASURED cells, permanently non-data-driven parameters, and one fully authored asset set that

cannot render.

A fifth thing binds nothing but costs everything: two instruments are lying. The foot-slide mean

and the drifted far-field station both produce plausible numbers about the wrong subject. Rows 8, 12

and 51 close that class, and they are cheap.

---

6. The waves this register proposes

The standing queue (HANDOFF section 8.7) is Q1-Q10 and is asset-wave shaped. Six gaps in it surfaced

while ranking. These are BEYOND the floor of what was asked and are named so Josh can veto any of

them; none of them reorders an existing Q wave except where stated.

---

7. Provenance

Six domain reads, each cold against its own records, synthesized here 2026-08-06. Spot-verified at

HEAD by the synthesizer rather than carried on trust: remesh_used=false / out_quads=0 /

topology_route=tri_decimate / uv_area_utilization=0.5577 in the face finish record;

admissible=false and positive_controls_responded=false in control_surface.json; 182 corpus rows

at 4 ACQUIRED / 178 MISSING against 132 formula cards on disk; T0_Boss_Attack_Rail at 27 columns

and 0 data rows; T0_Creature_Roster at 187 rows; the three garment LOD0 triangle counts in the

arrival-set manifest; and the existence of all eleven cited canon-side record files.

Every number in this register belongs to the lane that measured it. Where a lane declared a figure

UNMEASURED, INADMISSIBLE, or RETRACTED, this file carries that word rather than a substitute.

Generated by harness/site/structure_site.py — the URL path is the repo path. review root