PIPELINE_STACK_SYNTHESIS_2026-07-29.md

pipelines/PIPELINE_STACK_SYNTHESIS_2026-07-29.md

PIPELINE_STACK_SYNTHESIS — the eight lanes folded into one factory plan

Synthesized 2026-07-29 from PIPE_ART · PIPE_3D · PIPE_ANIMATION · PIPE_VIDEO · PIPE_AUDIO_MUSIC ·

PIPE_VOICE · PIPE_WEBSITE · PIPE_CONTROL_PLANE. This document folds and flags; it does not

re-adjudicate a lane's own verdict, and §5 deliberately leaves conflicts open for the director.

Evidence tiering is carried through from the source dossiers and never upgraded here:

[V] = a primary source was fetched by the owning lane · [E] = arithmetic from a published

param count · [U] = honest unknown, recorded as such. A size or claim never gains confidence by

being copied into this file.

---

1. THE CLAIMS TABLE — every search-AI name the user quoted

Two honesty notes on this table. (a) Rows 9-10 were the user's leads that no dossier covered;

they are answered here from the repo's own prior verification plus one live confirmation pass, and

both land as *real but not adoptable on this hardware* — which is why the 14B pick (D-4) stays open by

design rather than being closed by a search-AI name. (b) Every "wrong lane" verdict above was reached

independently by 3-4 separate dossiers, which is the strongest signal in the survey: **the search AI

conflated a motion model, a music model and a world model into the image/audio stack, and four

lanes caught it separately.**

---

2. THE CONSOLIDATED INSTALL MANIFEST

2.1 Thursday-night download list — deduped, by destination drive

Ruled layout: Gen5 4TB slot 1 (C:) = Windows + UE engine hot path + DDC · Gen4 4TB slot 3 (D:)

= models / ComfyUI / projects cache · 8TB SATA slot 4 (E:) = archive.

→ Gen4 (D:) — all model weights, all venvs, the WSL2 vhdx

→ 8TB SATA (E:) — archive-class pulls

→ Gen5 (C:) — npm/tooling only (npm_config_cache redirected to D:)

Deduplication actually performed: SeedVR2 (claimed twice: ART + VIDEO), ARDY+Kimodo (claimed twice:

3D + ANIMATION), Blender (3D + ANIMATION), the cu128 torch wheel set (7 lanes → one pip cache), and

the ComfyUI weight tree (4 lanes → unresolved, see C-1). Naive per-dossier summation gives

≈420-520GB; the deduped figure is ≈350-450GB.

The control plane's Thursday deliverable is 0 GB and is the most important item on the list: the

download manifest itself — one JSON row per artifact carrying url, bytes, sha256, dest,

licence, so an interrupted overnight pull is *verifiable* rather than re-downloaded, and **a row

with no bytes/sha256 is rejected at write time.** Every [E] and [U] above is a cell that must

be corrected with a measured value on the night — the runbook already carries three

UNSOURCED-ESTIMATE VRAM rows and this list must not add a fourth class of them.

2.2 Friday+ install order, with the D-1 containment assignment per lane

**D-1 (runbook:1507) offers WSL2+CUDA vs ComfyUI-on-Windows vs dual-boot, and its own recommendation

is the split: "WSL2 for official-path models + ComfyUI-Windows for the image lane."** Every dossier

converged on a *per-lane* answer rather than a machine-wide one:

THE HEADLINE THAT RESHAPES THE WEEK — the 3D lane does not install as published. Every 3D-gen repo

pins a pre-Blackwell CUDA (TRELLIS.2 torch 2.6.0 + cu124, Hunyuan3D-2.1 2.5.1+cu124, Direct3D-S2

2.5.1+cu121, ComfyUI-3D-Pack Windows prebuilds cu124/torch-2.5.1 with zero 50-series note),

while the RTX 5090 is sm_120 and requires ≥ PyTorch 2.7.0 + cu128. So the install plan for that

lane is not "follow the README" — it is "rebuild every CUDA extension against cu128"

(flash-attn, spconv, torch_scatter, torch_cluster, torchsparse, and TRELLIS.2's

cumesh/o-voxel/flexgemm). **This is a half-day-to-two-day item and it is the single biggest

unbudgeted cost in the whole factory setup.** The runbook's cu124 pin for TRELLIS.2 is simultaneously

CONFIRMED as the repo's published requirement and CONFIRMED as unrunnable on sm_120 — read it as

"the version the authors tested," never "the version we install."

Install order (ascending risk, so the box is productive before the hard lane starts):

• PHASE 0 — Thursday evening, on-box, zero GPU. Claude Code sign-in · HF token into the env ·

write the download manifest · Stage-6F firewall pre-authorization including 8188 AND 8001

(an unacknowledged Defender dialog at 04:00 presents as a hung editor, not an error — the worst

overnight signature in the house) · kick the Tier-1 pulls.

• PHASE 1 — Friday AM, GPU-FREE, before the first GPU hour. (a) The licence re-read pass —

ART, VIDEO, VOICE and CONTROL_PLANE each independently scheduled this as a hard pre-flight; fold

them into one sitting: HiDream-O1 MIT · Qwen Apache · Z-Image Apache · SeedVR2 Apache ·

Depth-Anything-3 per-variant · Gemma-4's apache-2.0-tag-vs-Gemma-licence-link discrepancy

(the entire HiDream text-encoder chain depends on it) · HY-World's unreadable License.txt ·

Kimodo per-variant · every upscale weight against the allowlist. (b) **The Batch-0 registry

backfill** — see §2.4; it is the head-of-chain for every lane's dispatch and costs zero GPU hours.

• PHASE 2 — the base env. Python + torch==2.9.0+cu128, then **verify sm_120 before anything

else**: python -c "import torch;print(torch.cuda.get_device_capability())" must return (12, 0).

Nothing downstream is trustworthy until that prints.

• PHASE 3 — lanes, in this order: (3a) WEBSITE — zero GPU, buildable pre-box today on the

current machine · (3b) CONTROL_PLANE — D:\pipe\.venv, requests websocket-client psutil,

lanes.json, routing_table.json, the two scheduled tasks · (3c) VOICE CPU sub-lane — 10 minutes,

zero GPU, and the first proof the box generates anything at all · (3d) ART — ComfyUI Windows

(pin a cu128 wheel; the portable README currently ships cu13.0 and D-2 says avoid 13.x) →

Manager → Z-Image first light (smallest/fastest/Apache) → HiDream-O1 → Depth-Anything-3 →

SeedVR2 → Krita · (3e) VOICE GPU sub-lane — TTS-Audio-Suite → Qwen3-TTS → Chatterbox

gate-locked clone_enabled=false by default · (3f) AUDIO — ACE-Step venv → Gradio → REST :8001

→ MOSS → Demucs · (3g) ANIMATION — Kimodo venv → SOMA ckpts → **write Tools/npz_to_bvh.py

ourselves** (~1 file, licence-clean, ENGINE-class, removes both unverified third-party bridges) →

Blender bake → UE import → *only then* ARDY as an eval lane · (3h) VIDEO — Wan 2.2 first (the

licence-clean baseline) → SeedVR2 → HunyuanVideo-1.5 (the co-residency probe) → LTX-2.3 last,

because it is the one that may not fit and failing there costs nothing · (3i) 3D LAST — WSL2 +

Puget pack + Container Toolkit → cu128 base → the extension rebuild → TRELLIS.2 geometry-only

(setup.sh WITHOUT --nvdiffrast --nvdiffrec) → Direct3D-S2 → Hunyuan3D-2.1 → the CPU finishing

set → UniRig → HY-World → ARDY/Kimodo.

*The ordering principle, stated because it inverts importance:* 3D is the most valuable lane and is

installed last, because it is the only one whose install can consume two days, and every other

lane is independently useful on day one. A factory that spends Friday on flash-attn has nothing

running Friday night.

2.3 The conflict pins — one answer per contested knob

2.4 The one blocker that precedes every lane's autonomy (zero GPU, Friday AM)

✅ CLOSED 2026-08-02 — BATCH 0 APPLIED, gates green (36/36, exit 0). The 2026-07-29 numbers

below are kept as the dated record; the tree had MOVED by the time the work ran, and the honest

re-count is: 863 rows carried generation_status and 826 were the empty string, with the

dispatcher predicate returning 23 work orders, not 21. The delta is real movement, not a

miscount — T0_SFX_Registry has since GAINED all four assetgen status columns (so the "no

generation_status column at all" line below is stale; +30 rows carrying the column, all

empty), and T0_Theme_Registry gained two cardinal rows at the motif-roster apply (12 → 14

pending). The 796-empty figure across the eight mesh-bearing registries was still exactly

right. One value outside the enum that this section did not count was also found and handled:

T0_Voice_Registry.GRAND_SAGE_REVEAL_VOICE held the free-text

excluded_non_lexical_HL_0046, so the true residue was 827, not 826.

What landed, all in one commit: 843 generation_status cells written through the sanctioned

writer harness/apply_generation_lifecycle_backfill.py (826 empty → pending, 16 on the

brand-new T0_Scene_Spec_Registry column populated in the same pass, 1 declared old-value-

guarded retag); 17 columns appended across 12 registries including the C-12 generation_licence_ref

everywhere; docs/generation_lifecycle.json as the contract; docs/licence_records/ +

RECORD_SCHEMA.json + harness/generation_prompt_hash.py as THE LICENCE SCHEMA COMMIT; and the

vacuity guard as GATE 36 harness/check_generation_lifecycle.py, which is stronger than this

section asked for — it measures THE RESIDUE (rows the dispatcher's predicate vocabulary cannot

name) and fails unless it is zero, so the defect class is unrepresentable rather than merely

fixed. The dispatcher predicate now returns 865 work orders. Ten live mutation tests plus 27

fixtures prove the teeth are armed.

Counted on the live tree: 831 rows carry generation_status and 796 of them are the EMPTY STRING

(Antagonist 64 · Boss 281 · Character 164 · Creature 150 · Equipment 1 · Familiar 22 · Vril_Site 42 ·

Weapon 72). Empty string is outside the ratified five-value enum, so a dispatcher scanning

WHERE generation_status == 'pending' returns **21 work orders across the entire mesh surface and

reports the queue clean** — the "a search that CANNOT match reports zero" class, reproduced inside the

control plane. Plus: T0_SFX_Registry (30 rows) has no generation_status column at all (and

SFX is DR-2 §6 Batch 2, the *handshake exit*), T0_Scene_Spec_Registry (16 rows) carries a prompt

hash with no status column, and T0_Theme_Registry lacks regeneration_trigger.

Batch 0 therefore backfills pending on all 796 rows through the sanctioned writer, adds the missing

assetgen columns to SFX + Scene_Spec, refreshes fidelity_baseline.json in the same commit, and

ships a vacuity guard: if total enrolled rows is 0, or a registry with assetgen columns matches 0

rows, the run exits non-zero. *A queue that cannot match reports zero and looks exactly like a

finished night.* Four lanes independently want schema changes in this same commit — see C-12.

---

3. THE CONCURRENCY BUDGET SHEET — proving the factory day fits 32GB / 128GB

Governing arithmetic, inherited verbatim: *"29GB of 32GB is the tight fit the whole hardware decision

turns on … heavy lanes run sequentially, not concurrently"* — and **an OOM caused by our own

co-residency sloppiness would be a FALSE trigger on HARDWARE_DECISION's upgrade trigger #1.** Every

lane below signed up to that rule independently.

3.1 The four scheduling classes

CLASS A — ALWAYS-ON · 0 GB VRAM. Never holds the lease; never contends.

CLASS B — WINDOWED · the daytime UE window. UE holds the lease; exactly ONE small tenant beside it.

CLASS C — OVERNIGHT / EXCLUSIVE. One job at a time, serialized by the queue, never co-resident.

CLASS D — HARD RESERVATION. The 04:00 soak (D-11b). Asset-gen never overlaps it. Not negotiable.

3.2 The factory day, proven against the clock

**Throughput: 3 heavy GPU jobs per night + a full day of small-tenant generation + an uninterrupted

CPU rail. Each Class-C run carries a hard wall-clock ceiling** so a runaway lane cannot eat the

next day's window, and every run is self-stopping with a 30-second RUN_STATUS.json heartbeat (a beat

older than 10 minutes with a live process opens a stall record, not a crash — because a firewall

dialog and a hung editor look identical from outside).

3.3 RAM (128 GB) — does it stack?

Both fit, and the second is *why* the art lane's 32-48 GB offload requirement is affordable: it only

ever peaks in a window where UE is not cooking. Scheduling note, not a conflict: ART's "hold

32-48 GB free" and 3D's "~32 GB for the CPU half" are the two largest RAM claims and they stack

safely only while a UE cook is not also running. If all three coincide the box is at ~130 GB and over.

3.4 Disk claims against the RULED layout

• Gen5 (C:) — UE hot path + DDC. Claimed by: UE, the pre-existing canon repo, and **WEBSITE's

C:\dev\humanity-site\ (⚠ C-2)**. Four dossiers state "nothing from this lane touches Gen5."

• Gen4 (D:) — every model weight, every venv, the WSL2 vhdx, D:\pipe\, ComfyUI models/.

Total ≈350-450 GB against 4 TB. Comfortable.

• 8TB SATA (E:) — ART boards + the Lane-R National Geographic corpus · 3D raw-GLB archive +

DEM/WorldCover GeoTIFFs (archive once — they are rate-limited, not merely large) ·

E:\anim_gen\ · video reel archive · Sonniss · E:\archive\vo_masters\ + vo_reference\ ·

E:\pipe_archive\ (rejected outputs **kept, never deleted — a rejected asset is the QA gate's

evidence**) · website capture masters + press-kit staging.

---

4. JOSH-MINIMUM CONSOLIDATED — the complete list, nothing else needs him

**Tagged [THU-EVE] = on-box or blocks the Thursday-night pulls · [REMOTE] = anytime, from

anywhere, no box needed · [LATER] = deferred by its own ruling.**

Blocking the setup:

1. [THU-EVE] Claude Code sign-in on the new box. The on-box session owns everything after it.

2. [THU-EVE] HuggingFace account + token, and per-repo gated acceptances. Anonymous fetches of

both FLUX.2-dev/LICENSE.md and the Krea org returned HTTP 401 today. Clicking "accept" is an

account action; agents never handle tokens. *Blocks any gated pull on the night.*

3. [THU-EVE] Fab download of the UE Game Animation Sample (500+ anims) — his Epic account.

*Skip the Fab "Kimodo Importer" — we write the exporter ourselves.*

4. [THU-EVE] Sonniss GDC bundle download — free, but the site is email/form-gated. One sitting,

or delegate.

5. [REMOTE, DO NOW — pre-arrival] The Puget burn-in / benchmark sheet ask to the rep. Requested

independently by four dossiers (ART, 3D, VIDEO, CONTROL_PLANE) — it is the only pre-arrival

thermal/power baseline for this exact machine, and the window-sizing arithmetic wants it.

Rulings only he can make (each already briefed with a recommendation, never a bare question):

6. [REMOTE] D-1 ratification — the 3D lane asks him to confirm WSL2+Puget as the generation host

with ComfyUI-Windows as the image lane. *Note: six other lanes answered D-1 for themselves and

three of those answers are per-lane, not machine-wide (C-9).*

7. [REMOTE] Ship-territory ruling on the Tencent licences (Hunyuan3D-2.1 + HunyuanVideo-1.5 both

exclude the EU, UK and South Korea). Two sub-questions: does the clause block *us using* the

model, block *distributing outputs into* those territories, or both. **Recommendation (control

plane): drop Hunyuan3D from the ship path** — the geometry leg is already covered by MIT models and

a worldwide Steam release is worth more than one alternate mesh backend.

8. [REMOTE] SMPL-X / Meshcapade decision — are the non-SMPL ARDY/Kimodo skeletons sufficient, or

is a Meshcapade commercial licence wanted? *Recommendation: non-SMPL skeletons; retarget joint

rotations onto our own rig and never ship SMPL-X mesh data.*

9. [REMOTE, CONDITIONAL] Llama-3-8B: HF gated acceptance + the anti-Llama-posture ruling —

needed only if ARDY is adopted. *Recommendation: Kimodo primary, ARDY eval-only, which retires

the question entirely.* (Blocks download row 31; see C-3/C-4.)

10. [REMOTE] OpenTopography — the free non-academic key is 50 calls / 24h; 69 regions × 3

datasets ≈ 207 calls ≈ 4+ days of API budget. Accept the spread, buy OpenTopography Plus, or

take the AWS bulk-COG route (no account, no limit, more engineering). A spend/schedule fork.

11. [REMOTE] Board delivery medium — recommendation: a published page for the eight

mythological realms, local files for the fairy realm (the most spoiler-dense artifact the

project will produce). It is a publishing decision, so his nod.

12. [REMOTE] May any public marketing surface ever carry a generated frame? LTX-2's Attachment A

§5 compels an intelligible machine-generated disclaimer on any surface that does.

Recommendation: no — every public frame stays a real capture, which is already the ruled

Capture lane and costs the video lane nothing.

13. [REMOTE] Neither commercial image licence — BFL self-serve tiers (Builder 10K img/mo →

Enterprise) and Krea (Enterprise above $1M revenue) are both purchasable and both only he can

sign. Recommendation: buy neither; the MIT/Apache tier is genuinely competitive and the

licence-drift surveillance cost is the expensive part, not the sticker.

14. [LATER] Ship-time watermark call — keep Resemble's Perth neural watermark in shipped

Chatterbox VO as a provenance asset, or strip it (MIT permits). *Recommendation: keep.*

Accounts and money:

15. [REMOTE] ElevenLabs account + payment — commercial licence starts at Starter $6/mo; the

15-node hero budget likely wants Creator ($11) or Pro ($99). He creates and pays; the key enters

the env like every other credential.

16. [REMOTE, OPTIONAL] Cascadeur Indie $96/yr (converts to perpetual after one year) — the one

paid animation seat. Optional second: Rokoko Studio cloud fallback (~$120/yr).

17. [REMOTE, DEFERRABLE] Krotos Dehumaniser 2 + Reformer Pro ≈ $798 — creature vocalisation,

the only purchase the audio lane needs, deferrable to the creature pass.

18. [REMOTE, OPTIONAL] BOOM Library buyout $99-199 — only if benchmark day shows Sonniss+MOSS

leaves gaps.

19. [REMOTE] Website W1 — authorize the Cloudflare↔GitHub connection for Workers Builds, once

(one OAuth authorization, no secret ever enters our repo — recommended over a long-lived API

token). W2 — create the humanity-site GitHub repo. W3 — the first custom-domain attach

authorization (DNS records and certs are then created automatically; no manual DNS typing).

20. [REMOTE] W4 — name the press-contact address to print. A decision, not a task. Email Routing

stays deferred and is not being re-offered.

21. [LATER] W5 go-live flip (already RULED: at first real captures) · W6 the "Humanity

Forgotten" TM search/filing + the defensive-variant re-check at go-live.

Ongoing / conditional:

22. [REMOTE] Lane-R rights. The National Geographic run is his physical property, not a

publication licence — NG plates are internal underlay only. **own_capture (his own photography)

is the cleanest rights class available and only he can create it**; likewise any per-site

commercial-capture permission.

23. [CONDITIONAL] Voice-talent consent records. If a real person is ever cloned, the §4.4.1

five-stage authorization is his signature, held by him. *The loop generates; it never authorizes.*

24. [ONGOING] Choice-board rulings when the realm art program surfaces boards — the one visual

artifact that legitimately reaches him before THE JOSH GATE, because a board is a *decision

surface*, not a playtest.

Explicitly NOT his: D-3/D-4/D-5/D-6 model picks (director-ruled on benchmark evidence) · lane

scheduling · model promote/demote inside the pinned roster · retries · the Batch-0 backfill · every

registry edit · all site content, CI and i18n · every QA gate.

---

5. CROSS-DOSSIER CONFLICTS — flagged, deliberately unresolved

These are places where two or more dossiers claim the same resource, home, or primacy differently.

I am not resolving them. Each names the claimants so the director can rule once.

C-1 · THE COMFYUI QUESTION — the largest structural conflict in the survey. How many ComfyUI

instances exist, on which side of the WSL2 boundary, and whose models/ tree do they share?

• ART: Windows-native, and explicitly declines the Puget pack ("stays the WSL2/3D lane's asset;

this lane does not need it").

• VIDEO: WSL2 + Puget comfy_ui as its host (option a), with Windows as a fast-iteration lane.
• VOICE: rides "the same ComfyUI the image lane already installs at Stage 6b" (i.e. ART's Windows one).
• AUDIO: drops ace_step_1.5_turbo_aio.safetensors into "ComfyUI/models/checkpoints/" — unspecified which.
• CONTROL_PLANE: models one ComfyUI HTTP+WS endpoint on port 8188.

Two instances cannot both hold 8188; SeedVR2 is claimed by ART and VIDEO simultaneously and would

be either duplicated across the boundary or shared through the vhdx. The download manifest, the

firewall table, and the dispatcher's lane map all assume a single answer that does not currently exist.

C-2 · GEN5 (C:) EXCLUSIVITY. ART ("C:/Gen5 stays UE-only"), 3D ("nothing from this lane touches

it"), VOICE ("no model weights"), and CONTROL_PLANE ("nothing from this lane") all assert Gen5 is

UE-only. WEBSITE places C:\dev\humanity-site\ on Gen5 by design, arguing it is a git-hot,

latency-bound small working tree and not a model store — and separately that it must NOT live inside

C:\dev\humanity-forgotten (a node_modules inside the gate-enrolled canon tree collides with

run_gates.py and the fidelity baseline). Note the canon repo already lives on C:, so "UE-only" is

already inexact. Two sub-questions ride this: whether website builds on Gen5 are an acceptable

neighbour for DDC/shader-compile I/O, and whether the "must not run during a UE cook" rule is

sufficient mitigation.

C-3 · ARDY vs KIMODO — three answers on primacy, two on lane ownership.

• 3D §1 motion row: ARDY PRIMARY, Kimodo FALLBACK, inside the 3D lane.
• ANIMATION: Kimodo PRIMARY, ARDY EVAL-ONLY (because of the gated Llama-3-8B encoder and the

repo's standing anti-Llama posture), owning the lane outright.

• CONTROL_PLANE routing table: ARDY pinned for "character motion / anim", Kimodo for

"constrained/keyframed" — a *split by job*, which is a third shape.

• ART: "hand them to the animation lane, not here."

The routing table is the committed artifact, so whichever way this lands must be written there.

C-4 · THE LLAMA-3-8B RIDER IS INVISIBLE TO TWO OF THE THREE CLAIMANTS. ANIMATION names ARDY's

text encoder as gated meta-llama/Meta-Llama-3-8B-Instruct (~16 GB, Llama 3 Community License, a

Josh HF acceptance, and a live conflict with the repo's anti-Llama posture). 3D lists ARDY as motion

primary describing only a "text encoder ~14GB VRAM (CPU-offloadable)" without naming Llama or the

gate. CONTROL_PLANE pins ARDY with no rider at all. **If ARDY is primary, Josh-minimum gains an item

that two of the three dossiers do not know about** — and a 16 GB gated download joins the night list.

C-5 · HUNYUAN3D-2.1 IS SIMULTANEOUSLY A SHIP PRIMARY AND A PREVIZ-ONLY ROW. 3D lists it as the

"Finished textured PBR | PRIMARY." CONTROL_PLANE's routing table marks it **previz-only until Josh

rules on ship territory**. VIDEO applies the same §2.3 posture to its sibling HunyuanVideo-1.5 and

concludes "internal previz only, permanently." All three cite the same clause (1M MAU + EU/UK/KR

exclusion). It is the only model in the survey occupying a ship-path slot under a licence the ruled

posture classifies previz-only — and it is also the 29 GB ceiling case the whole concurrency

budget is sized around.

C-6 · COSMOS 3's LICENCE-READ STATE DIVERGES. VIDEO fetched nvidia/Cosmos3-Nano and read

OpenMDW-1.1 ("ready for commercial and non-commercial use" — the cleanest licence in the survey).

CONTROL_PLANE's routing row records "not read." 3D excludes it on capability grounds without a

licence read. The *verdict* agrees everywhere (excluded), but the routing table would ship carrying a

stale "not read" against a licence that has in fact been read — and VIDEO explicitly flags Cosmos as a

WATCH item that would jump to public-surface-eligible if ComfyUI support lands.

C-7 · SEEDVR2 HAS TWO OWNERS AND TWO SIZES. ART wants SeedVR2-3B NVFP4 (~3 GB) in the Windows

ComfyUI as its upscale/finish step; VIDEO wants SeedVR2-3B (~7 GB, +7B optional ~15 GB) in the

WSL2 ComfyUI for 720p→delivery. Same Apache-2.0 model, two hosts, two download rows, two quantization

choices. Downstream of C-1.

C-8 · "EXCEL BASS" CARRIES THREE DIFFERENT VERDICT TOKENS. AUDIO: "CONFIRMED as XL BASE."

CONTROL_PLANE: "KILLED as a name" (a transcription artifact of xl-base). VOICE:

"UNVERIFIED-EXCLUDED" (correct within the voice lane — no such TTS model exists). The underlying

facts are compatible, but a config or registry cell will inherit whichever token it read first. One

canonical verdict string is needed: *the name is refuted, the referent acestep-v15-xl-base is real.*

C-9 · D-1 HAS NO SINGLE ANSWER, AND ONE LANE'S RATIFICATION ASK IS PHRASED MACHINE-WIDE. 3D asks

Josh to ratify "(a) WSL2+Puget as the generation host and (b) ComfyUI-on-Windows as the image lane

only." But AUDIO declares it has "removed the audio lane from the D-1 fork entirely," ANIMATION

declares it "must NOT be folded into the ComfyUI install," VIDEO says (b) is "genuinely viable here,"

and WEBSITE + CONTROL_PLANE say D-1 does not bind them at all. A per-lane D-1 may be the right answer

— but the ratification currently on offer would read as a machine-wide ruling that four lanes have

already departed from.

C-10 · FIREWALL AND PORT OWNERSHIP GAP. AUDIO flags that **port 8001 (ACE-Step REST) is missing

from the Stage-6F firewall table**, which lists UE, MCP 9315, MCP 8000, Ollama 11434 and ComfyUI 8188.

CONTROL_PLANE owns Stage-6F pre-authorization and does not list 8001. WEBSITE declares no dev-server

port at all. The consequence is named by the control plane itself: *an unacknowledged Defender dialog

at 04:00 presents as a hung editor, not an error.*

C-11 · BLENDER: ONE BINARY, TWO OWNERS, TWO SCHEDULING CLASSES. 3D installs Blender 5.2 LTS

(Python 3.13) as its CPU finishing host and classes it ALWAYS-ON (the lane's throughput win).

ANIMATION installs "Blender 5.x" as its headless FBX bake host and puts that bake **in the overnight

window**. Sharing the binary is fine; the two scheduling-class claims on the same CPU budget are not

reconciled, and both lanes independently budget ~8 threads.

**C-12 · FOUR LANES WANT A SCHEMA CHANGE IN THE SAME BATCH-0 / PRE-FREEZE COMMIT, UNDER THREE NAMES

FOR THE SAME FACT. VOICE asks for generating_model + model_licence** columns on voice rows.

ANIMATION asks for asset_licence_ref on the assetgen block at Batch 0. CONTROL_PLANE declares

model_licence + previz_only mandatory in a git-tracked *artifact record beside the output*

(not a registry column) plus the 796-row generation_status backfill plus new status columns

on T0_SFX_Registry and T0_Scene_Spec_Registry. AUDIO independently asks for regeneration_trigger

on T0_Theme_Registry and all four assetgen columns on T0_SFX_Registry. ART adds a fifth shape: a

licence-class token folded into generation_prompt_hash so a model's demotion fires

regeneration_required automatically. All five are the same requirement — *name the generating

model's licence per artifact* — and they would collide in one commit with one baseline re-emit.

C-13 · THE "TWELFTH DR-2 CLASS" ORDINAL. ANIMATION proposes a twelfth §2 row

(Animation / montage, keyed on ability_id + presentation_tier_ref). VIDEO explicitly refuses to

invent one ("Video generation is NOT a DR-2 asset class, and I am not inventing a twelfth one") and

routes itself through the existing Cinematic-scene row as reference material only. Both positions can

be simultaneously correct — but they would race for the same ordinal in the same schema commit as C-12,

and ANIMATION's row depends on an unruled PC-5 decision (that the canon row is authoritative and

the montage is retimed to it) without which "a generated 1.4s windup silently rewrites a 0.9s canon

row, 200 times."

C-14 · TWO LANES ASSERT THE SAME LARGE RAM RESERVATION. ART requires 32-48 GB held free for

text-encoder CPU offload; 3D budgets ~32 GB for its always-on CPU finishing half. Both are correct

in isolation and both fit — *unless* a UE cook (32-48 GB) coincides with both, which puts the box at

~130 GB of 128. Not a contradiction; an unscheduled overlap nobody owns.

---

6. WHAT THE SURVEY AGREES ON (recorded because unanimity is itself evidence)

Eight independently-researched lanes converged, without coordination, on all of the following:

the 29-of-32 GB serialization law and that a self-inflicted OOM is a *false* hardware-upgrade

trigger · licence-read-before-first-GPU-hour, with the operative *document* re-fetched rather than

a model card summarized · the generating model's licence named in every artifact, mechanized so a

demotion fires regeneration_required rather than depending on someone remembering · **dispatch from

script-emitted registry rows, never a hand-typed worklist · care-gated ids never raw-generate** —

the composer refuses on a column · previz-tagged outputs can never promote · **a tooth without a

failing fixture is not armed · and no quality claim without a capture behind it**, read by a fresh

critic, with the director re-reading the deciding frames from the newest session before any claim

enters a commit message.

Three lanes are zero-GPU and can start before the box arrives: WEBSITE (fully buildable today on

the current machine — Node 24.15.0 confirmed present), CONTROL_PLANE (0 GB, ~2 GB RAM), and the

Batch-0 registry backfill that unblocks every other lane's autonomy.

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