pipelines/STACK_FACTS_QUESTIONS.md
The correction pass (Phase D) needs ground truth Fable cannot derive. Eight questions, each with
the survey evidence and a RECOMMENDATION per the decision protocol — answer inline, adjust
freely; everything else in the review proceeds without waiting on these.
OWNED: Josh already bought NeoStack (lifetime). UE 5.8: CORRECTED (Q3 refresh, 2026-07-15) —
the Discord quote names SIK = the STEAM/EOS Integration Kit (a different Betide product), NOT
the Agent Integration Kit (NeoStack); every public NeoStack source still caps at UE 5.7, so
NeoStack-on-5.8 is UNVERIFIED-from-public-sources. Josh's own hands-on ("ue5.8 I think worked
with neostack") stands as the operative datapoint — the first-use verification (already a listed
spike) settles it. Development active regardless (V2 stable, V3 + 4 products incl. their IDE
announced). Their bundled gen RESOLVED: 3D = a Meshy+Tripo cloud passthrough; voice =
ElevenLabs — complements, does not replace, local TRELLIS-class gen. The v2 announcement:
covers the ENTIRE engine, every asset type/workflow, works with Claude natively; the lifetime
purchase bundles weekly usage incl. 4-5 agentic models, 3D ASSET GENERATION, and VOICE
GENERATION (a new datapoint: NeoStack's own gen lanes join the Q3/Q4 evaluation set).
THE RULED ARCHITECTURE PRINCIPLE (Josh, verbatim intent): Claude Code drives NeoStack, and
every action — code, Blueprints, procedural generation, LLM-placed generation — is MATCHED TO
OUR SCHEMA (the v3 spine graph + region pages + registries are the input contract; no freehand
generation). The UE Translation Doc is written to that principle; the only residual check is
routine per-plugin verification at first use.
Josh's screenshots settle the referent: Cosmos 3, launched May 31 2026 — "the world's first
fully open omnimodel," native reasoning + WORLD AND ACTION GENERATION, Mixture-of-Transformers,
downloadable; Google's AI Overview itself now calls it "the NVIDIA Open World Generator." The
first research pass dismissed Cosmos on Cosmos-1/2-era reads — that dismissal is WITHDRAWN
pending the dedicated deep dive (tech_research/COSMOS3_DEEP_DIVE.md, running now), which
re-derives the game-dev applicability from primary sources adversarially both ways, including
the output→UE bridge toolchains (a parallel bridges lane) and the world-models landscape
(a third lane). The paragraph below is the SUPERSEDED first-pass verdict, kept for provenance:
The research lane (tech_research/NVIDIA_WORLDGEN.md, fully cited) found NO NVIDIA product that
is an open-world generator in our sense (geodata in → populated game world out). What actually
exists as of mid-2026: (a) NVIDIA's one shipped, confirmed-5090-local generative-3D piece is the
AI Blueprint for 3D Object Generation (TRELLIS NIM) — text → Blender-ready GLB PROPS, an
object-layer tool that slots into prop population, not terrain; (b) NVIDIA's 2026 push (Cosmos 3,
Earth-2, GTC Omniverse) is robotics/climate infrastructure — no meshes, no game worlds; (c) the
"describe it, get a world" products demoed at GTC 2026 are NON-NVIDIA (Overworld — a VC-funded
local diffusion world model; Promethean AI; Tencent's OpenUSD pipeline); (d) UE 5.8 itself ships
a new native Mesh Terrain system (Epic's, the closest 2026 thing to the slot); (e) the Omniverse
UE connector caps at UE 5.3 — on our 5.8 target, Epic's native USD plugin is the only
interchange regardless.
QUESTION: what did you have in mind when you said "the nvidia open world generator" — the GTC
open-world demos (likely Overworld, which is not NVIDIA), the TRELLIS object-gen Blueprint, or
something else you saw?
RECOMMENDATION: keep the researched terrain chain (Copernicus DEM → GDAL alteration → UE
Landscape/Mesh-Terrain, per tech_research/GEODATA_TERRAIN_CHAIN.md — free, scriptable,
commercial-clean); wire TRELLIS-class local gen in as the PROP-POPULATION step alongside UE PCG;
evaluate Overworld separately when it ships publicly (it is the actual "open-world generator"
shape); retire the nvidia_worldgen terrain-enum premise.
"The best open source that runs on my 5090 that beats Meshy and the others" — is that a
pick-at-build-time directive (we benchmark when the workstation lands) or do you already have a
candidate (TRELLIS / Hunyuan3D-2 class)? And does it take ALL seVEN mesh categories (weapons,
creatures, familiars, NPCs, vril sites, architecture, special-asset-class) or do some (MetaHuman
NPCs) stay engine-native?
RESEARCHED (tech_research/LOCAL_3D_ASSET_GEN.md, license files verified directly): split by
task, not blind preference — TRELLIS.2 (MIT) for geometry/blockout BUT its PBR texture-bake path
depends on NVIDIA nvdiffrast/nvdiffrec, which are NON-COMMERCIAL for third parties (a real
ship-blocker on that leg); Hunyuan3D-2.1 for finished textured PBR BUT its license caps 1M MAU
and EXCLUDES EU/UK/South Korea (a territory question for a shipped game — needs your read);
Hi3DGen (MIT) as the sharpest-geometry fallback; avoid Sparc3D (no license, no weights) and the
closed-API-only tiers. The 29GB-VRAM finding: your 5090 (32GB) is the first consumer card that
runs Hunyuan's full pipeline. NPCs stay MetaHuman-first; REAL CULTURAL ARCHITECTURE should not
be raw-generated at all (no accuracy guarantee — a §17.1 care call); the retopo/LOD/UV gap has
no mature open AI solution — Blender-headless scripting + human QC. Section 5 of the brief is
the ready-to-run 5090 benchmark protocol.
JOSH-RULED (2026-07-15): licensing never constrains where the game ships or Josh's ownership —
the game markets to the US and everywhere else; the architecture and story are Josh's, the AI is
a tool. OPERATIONALIZED (this is the ruling's teeth, not a hedge): the pipeline DEFAULTS to
commercially-unrestricted components (MIT-class: TRELLIS.2 geometry, Hi3DGen, a clean texture
leg picked at the benchmark) so that NO tool's terms have any claim on the game anywhere —
the flagged restrictions were tool-EULA terms (the same class as engine royalty terms), never a
question of authorship, and the clean-toolchain default makes the ship-everywhere goal
unconditional. Territory-restricted tools (Hunyuan) are simply not load-bearing: usable for
internal previz/reference at will, in the shipped-asset path only if their terms someday clear.
NeoStack's own bundled 3D-asset gen joins the benchmark set.
AIVA (music, composer-in-loop), ElevenLabs (voice, with the §4.4.1 clone-authorization gate),
and the SFX chain (ElevenLabs SFX endpoint, Stable Audio Open, Boom/Pro-Sound-Effects/Soundly
libraries) are NOT covered by the three named replacements. Staying or going?
RESEARCH-CLOSED (2026-07-15 — tech_research/AUDIO_STACK.md, live-verified per tool; supersedes
the earlier keep-as-written placeholder). The evidence-backed lane recommendations now standing:
SUPERSEDED 2026-07-29 (the AIVA sweep, ledger 28ai): AIVA is CLOSED-AS-SKIP (Josh, 2026-07-27 — the enterprise-contact path is dead; nobody emails anyone). Bulk music = ACE-Step/YuE local, benchmark-gated; the hero/canonical theme is composer-in-loop; ACE-Step 1.5 ships native stem separation, retiring the only-AIVA-exports-stems premise below. Current stack: docs/pipeline_review/tech_research/PIPE_AUDIO_MUSIC_2026-07-29.md. The text below is preserved as the pre-ruling research record.
Pro tier (~€33/mo annual) grants copyright-owned-by-you, and AIVA has NO self-serve public API
(contract §3.4's orchestration-agent REST assumption; API terms need a negotiated agreement —
contact to settle). AIVA is also the only vendor with native per-instrument stems (the adaptive-
layering requirement). Bulk regional themes: ACE-Step 1.5 (MIT) + YuE (Apache 2.0) — both
UNRESTRICTED, both 5090-benchmark-gated. Udio is DEAD for our use (no file export at all under
the UMG walled-garden deal); Suno carries reverse-indemnification + active UMG/Sony litigation;
Meta MusicGen/AudioCraft weights are CC-BY-NC (non-commercial) — none load-bearing, ever.
fill (MOSS-SoundEffect v2.0, Apache; Stable Audio Open is revenue-capped so previz-class) +
Krotos Dehumaniser/Reformer for the 132-creature vocalization lane (sound-designer-in-loop, as
the contract already frames it) + UE MetaSounds procedural for UI.
JOSH CONFIRMS the tiering (non-blocking); the audio Translation Doc + the §3/§4/§10 contract
corrections are written to these findings.
Already superseded in-repo (no question needed): ChromaDB → sqlite-vec + fastembed (built,
real); Ollama → optional (fastembed primary). Open: (a) the OCR leg — the docs say
ocrmypdf/tesseract (not installed yet); keep that plan, or another OCR route? (b) anything else
you know you're dropping that the survey didn't flag?
RECOMMENDATION: standardize on ocrmypdf+tesseract (install now — it is the book-processing
critical-path blocker), fastembed embeddings, sqlite-vec store; retire every ChromaDB/Ollama
mention in the correction pass.
The contract carries TWO different Drive paths for region pages plus a third canon/region_pages/
proposal; the real scaffold is _source/02_Tier_2_Region_Pages/. Ratify: the 69 pages live at
_source/02_Tier_2_Region_Pages/<region_id>.md?
RECOMMENDATION: yes — the scaffold you created is the home; the correction pass re-points every
path to it.
Live registries use snake_case slugs; the contract says T8_REGION_NNN_REGION_NAME; the Schema
Dictionary says REG_NNN. One ruling.
RECOMMENDATION: the live snake_case slugs (human-readable, already in the data, zero migration);
the template abstracts region_id behind one resolver either way, so a later re-key stays cheap.
The workstation purchase is gated on it. Options: (a) STRUCTURE-PROVEN — OCR installed, the
pilot re-run + persisted (manifest record + registry rows + processed/ artifact), the auto-sort
gazetteer expanded, the special-format recipes each proven on one real item; (b) INBOX-DRAINED —
(a) plus all 658 staged files + the sacred-texts DVD sorted/ingested/registered.
RECOMMENDATION: (a) as the purchase gate (structure-proven means the 5090 lands into a working
pipeline it can immediately accelerate), with (b) as the running-in-parallel bulk work the local
tier chews through before and after the workstation arrives.