pipelines/ADVERSARIAL_ADJUDICATION.md
Fresh-context adversarial pass over docs/pipeline_review/tech_research/* + the three consolidations
(RESEARCHED_STACK.md, STACK_FACTS_QUESTIONS.md, GAP_MAP.md). Job: attack the body of work
before the owner builds on it. Read-only on the repo; no git commit. Research/verification date
2026-07-15. Load-bearing contested claims were re-checked against primary sources this session
(fetches cited inline); everything else is analysis of the corpus as written.
Meta-finding, stated up front because it changes how to read everything below: the nine briefs
are unusually disciplined — VERIFIED/INFERRED tagging is honest, and every primary source I
re-fetched this session (Lyra-2.0 card, Cosmos3-Nano card, NeoStack/Fab, ACE SDK) **resolved and
matched the brief's underlying facts.** The corpus is NOT built on hallucinated citations. The
errors that exist are almost all one of two kinds: (a) a license mis-read (code license quoted
as if it covered weights), and (b) the **consolidation layer asserting firmer claims than its own
briefs support.** Both are catch-at-review-time errors, not fabrication — which is the good news and
also exactly why this pass matters, because the two loudest errors are wired into the top-level
RESEARCHED_STACK.md/STACK_FACTS_QUESTIONS.md that Phase D would write against.
---
| # | Item | Verdict |
|---|---|---|
| 1a | Lyra 2.0 license | RULED: COSMOS3 correct — weights are NON-COMMERCIAL research-only. VIDEO_TO_3D_BRIDGES is WRONG (Apache-2.0-both claim). |
| 1b | NeoStack UE 5.8 | RULED: briefs correct (UNCONFIRMED). RESEARCHED_STACK.md "5.8 CONFIRMED SHIPPING" is UNSUPPORTED — overstatement. |
| 1c | Cosmos 3 model matrix | RULED: 16B Nano / 64B Super (verified). NVIDIA_WORLDGEN's table (2B/14B) is stale Predict2.5 figures — uncorrected in-place. |
| 1c | HY-World 2.0 / Gaea 3.0 / splats | Briefs honestly hedged; no contradiction to rule, open items remain. One loose "first-party" label on Luma's UE plugin. |
| 2 | Assembled chains | Terrain chain CONFIRMED with caveats; PBR-texture leg AT-RISK; ws.\* save + boss-AI BET (original builds); MetaHuman-at-scale + 69-region throughput UNMODELED. |
| 3 | Care surfaces | Photogrammetry-costume rule and raw-architecture rule are CORRECT and load-bearing; one gap (splat-of-real-capture still needs a rights/consent gate). |
| 4 | Ruled constraints | Constraint (i) "unrestricted-by-default" has a real hole at the PBR-texture leg — no confirmed clean automated open component exists. ACE weights ≠ code license. EULA "prompt-based input" ambiguity touches constraint (ii). |
| 5 | Missing lanes | Animation/motion-gen, water/ocean, weather/sky, crowds, localization, foliage source-assets, Chaos/destruction — none researched. |
---
The contradiction: COSMOS3_DEEP_DIVE.md §3 says the released weights are under the "NVIDIA
Internal Scientific Research and Development Model License," explicitly non-commercial, with the
code separately Apache 2.0. VIDEO_TO_3D_BRIDGES.md §1.4.2 says "**License: Apache 2.0, code and
weights** … the cleanest commercial license found anywhere in this entire research pass, no MAU caps,
no territory exclusions." Both cite the same URL (huggingface.co/nvidia/Lyra-2.0).
RULING — COSMOS3_DEEP_DIVE is correct; VIDEO_TO_3D_BRIDGES is factually WRONG. I fetched the
model card this session. Verbatim from the card: *"The Model and any Derivative Model may not be
distributed, deployed, sublicensed, publicly displayed, publicly performed… You may not use the
Model or a Derivative Model in a production environment or for the purpose of generating works for
sale or distribution."* License name on the card: **"NVIDIA Internal Scientific Research and
Development Model License." The code is Apache 2.0; the weights are non-commercial
research-only.** (Source: fetched huggingface.co/nvidia/Lyra-2.0, 2026-07-15; corroborated by
search — "the pre-trained model weights are restricted to research and development purposes only, not
commercial deployment.") 14B params confirmed.
Why this is HIGH severity, not a footnote: VIDEO_TO_3D_BRIDGES then promotes Lyra 2.0 as a
"validated benchmark-gate candidate" whose selling point is being "genuinely open and commercially
clean… better licensing than either of LOCAL_3D_ASSET_GEN's two lead candidates." That entire
recommendation rests on the false reading. Under the JOSH-RULED "commercially-unrestricted-by-default"
posture, Lyra 2.0 weights are the opposite of unrestricted — production and
generate-for-distribution are expressly forbidden. **Correct placement: Lyra 2.0 is
previz/reference-only, the exact same tier as Hunyuan3D and Marble — never in the shipped-asset
path.** COSMOS3 already placed it there correctly; the corpus contradicts itself, and the wrong side
is the one a reader reaches first in the bridges brief. The 43GB-VRAM-floor question is moot for
shipping — the license kills the shipped path before hardware does.
The contradiction: RESEARCHED_STACK.md L52–54 states NeoStack "**5.8 support CONFIRMED SHIPPING
per the developer** (Fab builds current; V2 stable, V3 announced)." Both NEOSTACK_AI.md (§0.3, §5)
and Q3_2026_MODELS_REFRESH.md (§2.5) say the opposite: no public source confirms 5.8; every
NeoStack/AIK source caps at UE 5.5–5.7; the Discord quote Josh has on record names SIK/EIK
(Steam/EOS Integration Kits — different Betide products) not AIK (Agent Integration Kit =
NeoStack).
RULING — the briefs are correct; RESEARCHED_STACK.md overstates. I confirmed this session that
betide.studio, neostack.dev, and the Fab listing still state UE 5.5/5.6/5.7 with no 5.8
tag; the "v1 is out now with 100% engine coverage" marketing line is a v1.0.x claim, not a 5.8
confirmation. The "V2 stable, V3 announced" clause in RESEARCHED_STACK.md is also unsupported —
Q3_2026_MODELS_REFRESH.md §2.4 found the official changelog "tops out at v1.0.50 (2026-03-24)" and
the lone "v2.0.45" datapoint is "uncorroborated… possibly confusing NeoStack with one of Betide's
other 8 products or simply wrong." Notably, Q3 §DELTAS already flags that STACK_FACTS_QUESTIONS
Q1's "CONFIRMED SHIPPING" line "NEEDS A RE-READ" — yet RESEARCHED_STACK.md, written the same day,
was never reconciled to that correction. **Action: strike "CONFIRMED SHIPPING" and "V2 stable, V3
announced" from RESEARCHED_STACK.md; the NeoStack-on-5.8 first-use spike is a genuine gate, not a
formality.** Josh's hands-on "ue5.8 I think worked" is a soft datapoint, not a substitute for the spike.
The discrepancy: NVIDIA_WORLDGEN.md §1 table lists "Cosmos 3 … 2B variant: local on one RTX
5090 (image-only); 14B variant: 64GB+ VRAM." COSMOS3_DEEP_DIVE.md §1 and WORLD_MODELS_LANDSCAPE.md
§2.A both say Cosmos 3 ships as Nano 16B / Super 64B, with Edge 2B announced-not-released, and the
2B/14B figures belong to the prior-gen Cosmos-Predict2.5.
**RULING — 16B Nano / 64B Super is correct; NVIDIA_WORLDGEN's figure is stale and uncorrected
in-place. I fetched huggingface.co/nvidia/Cosmos3-Nano: 16B parameters, OpenMDW-1.1 (commercial
5090.** COSMOS3_DEEP_DIVE self-flagged NVIDIA_WORLDGEN's error in prose, which is good practice — but
the WORLDGEN table itself was never edited, so a reader who opens WORLDGEN first still gets the wrong
number. Low functional impact (Cosmos is retired from the terrain premise regardless), but it is an
uncorrected factual error sitting in a brief. **The three briefs agree on the load-bearing point: Cosmos
3 is wrong output type (video/action, not geometry) AND a tight-to-infeasible 5090 fit unquantized.**
WORLD_MODELS_LANDSCAPE.md §2.B): correctly identified as tencent-hy-world-2.0-community, with EU/UK/KR-exclusion + MAU-gate **assumed by sibling-product
pattern, not confirmed text.** This is honestly hedged and matches the Hunyuan3D-2.1 pattern
LOCAL_3D_ASSET_GEN.md verified directly. No contradiction to rule; it stays an **open LICENSE-read
item** and, by pattern, previz-only — do not let it drift into the shipped path on the strength of
"open-sourced."
GEODATA_TERRAIN_CHAIN.md §6): "3.0 announced, ship-status unconfirmed"and "CLI-automation tier-gating unconfirmed (one source says Pro/Enterprise required; the
edition-comparison page doesn't mention it)." Honestly flagged; RESEARCHED_STACK.md reflects it
faithfully ("Open: Gaea CLI-automation tier + 3.0 status"). No overstatement here. **But note the
latent cost trap:** if headless CLI automation turns out to be Pro/Enterprise-gated, the "$99 Indie"
figure the consolidation leans on is wrong for the pipeline's actual (automated) use — confirm with
QuadSpinner before pinning $99.
UE 5.8 GS support; all plugins third-party; splats have no collision and bypass Lumen.** One loose
edge: COSMOS3 §3 calls the Luma AI plugin "the production-relevant first-party path" — Luma is
third-party, and VIDEO_TO_3D lists it correctly as such. Minor mislabel, worth a fix so nobody treats
a Luma plugin as Epic-blessed.
---
RESEARCHED_STACK.md)NASADEM / AW3D30 are genuinely worldwide + commercial-clean (attribution + a modified-data notice
only). Cesium-for-Unreal 2.28.0 confirmed on 5.8. PCG Production-Ready since 5.7 with first-class
DataTable ingestion. These rest on multiple independent Epic/vendor primary sources across two
briefs — solid.
-run=pythonscript, RunUAT BuildCookRun, PCGWorldPartitionBuilder headless bake, Horde-as-precedent — well-sourced
from Epic docs. Real.
here and the consolidation quietly picked the right side without saying so. GEODATA_TERRAIN_CHAIN.md
§3/Step-4 presents landscape_import_heightmap_from_render_target() as the headless "land it in
Unreal" step — but UE_BUILD_AUTOMATION.md §2.3 shows that stock Python only yields a
LandscapePlaceholder, LandscapeSubsystem exposes no one-call heightmap import, and the
render-target path imports into an already-existing landscape at render-target precision, not
full-fidelity creation. RESEARCHED_STACK.md correctly sided with "C++ shim wrapping
ALandscape::Import()," so the consolidation is right — but **GEODATA's own step-by-step is the
weaker read and should not be executed literally.**
UE_BUILD_AUTOMATION: (1) the Landscape import shim, plus (2) one C++ USTRUCT per registry (Blueprint-only structs
rejected — Epic bug UE-152346; ~29 registries, auto-generatable but still a codegen+compile layer),
plus (3) the ws.\* save subsystem (original build), plus (4) boss-phase→beat fight AI
(original build). Individually small; collectively this is the actual engineering-spike cluster on
the critical path before the Ch 2–5 slice. Calling it "a small C++ shim" undersells it.
UE_BUILD_AUTOMATION.md #9/§7.6 is explicit: **no Epic-documented many-flag worldstate pattern
exists; Lyra ships no save reference at all.** The TMap<FGameplayTag, FWorldStateValue> +
USaveGame design is synthesized best-practice, not a cited pattern. This is honest, but it means
two of the pipeline's load-bearing runtime systems are original engineering validated only by a spike
that hasn't run. Fine to proceed — but do not let "designed in the brief" read as "de-risked."
GEODATA_TERRAIN_CHAIN.md calls the GDALleg "deterministic, diffable, re-runnable" and batchable, but **gives no per-region wall-clock, no
disk-footprint, no total-batch estimate** for 69 regions × (DEM download + reproject + mosaic + clip
on is explicitly a deferred Josh design decision (§7), and Gaea erosion time per region is unbenched.
"The Ch 2–5 slice starts as VALIDATION of a specified chain" (RESEARCHED_STACK closing line) is true
for correctness, not for throughput — the 69× cost is unmodelled.
UE_BUILD_AUTOMATION.md §5.1/§5.2/§5.6:character editing is local since 5.6, but **auto-rig + texture synthesis still round-trip to Epic's
CLOUD per character (request_auto_rigging, request_texture_sources), and "cloud-service
throughput/rate-limits/latency at hundreds-of-characters scale is not documented."** For a game with
hundreds of named NPCs across 69 regions this is a genuine per-character serialized cloud dependency
with an unknown ceiling — plus MetaHuman Crowd (the masses tier) is Experimental with a "keep a
classic modular fallback" caveat. RESEARCHED_STACK.md says "spike it," which is right; flagging that
this is the most likely place the throughput math breaks contact.
NEOSTACK_AI.md §8–9: MILESTONE_BUILD_PIPELINE.md attributes Blueprint/Niagara/Sequencer/MetaHuman
authoring to "the UE 5.8 MCP Server plugin" — but Epic's actual native 5.8 MCP plugin ships only
Scene/Actor/Material/Object toolsets (no Blueprint, no PCG, no Sequencer, no Niagara). An autonomous
agent parsing that doc literally would execute against the wrong tool and hit a toolset gap it
doesn't expect. Must be disambiguated before any autonomous build loop reads it.
Nothing in the chains is dead-on-arrival. The failure mode is not "a step is impossible" but "steps
labelled settled are bets, and one cost axis (throughput × 69 regions × hundreds of NPCs) is
unmodelled." The chains are a specified hypothesis, correctly — just don't mistake specification
for validation, and don't mistake the tidy step list for a costed plan.
---
VIDEO_TO_3D_BRIDGES.md (§9 close) draws the right line: reconstruction tooling is cleared for
genuine captures of real places, but running AI-generated video of a real sacred/cultural
site through a splat/reconstruction pipeline produces an asset "wearing photogrammetry's credibility
as a costume" — all the visual trust signals of a survey, none of the ground truth. This is a sharper
version of the §17.1 risk, correctly identified as not a mitigation of it. The mitigation posture
is right: real-capture → reconstruction is the care-safe path; generated-video → reconstruction is
not. This directly closes the tool-gap LOCAL_3D_ASSET_GEN.md §3.4 left open (it flagged the need
but named no tool).
LOCAL_3D_ASSET_GEN.md §3.4: object-scaletraining data + no structural grounding + Stability's own "not factual representations" disclaimer
= never raw-generate a real sacred/cultural building; blockout-only, finish from reference/
photogrammetry/licensed kit. Both a quality and a care argument, correctly framed as a
recommendation to the art-pipeline owner, not a canon ruling.
The care briefs clear "real photos/video of the real place" as safe, and name sources including
"Josh's own travel captures, Wikimedia, licensed stock, tourist photos, drone footage." But
photogrammetry-from-others'-photos carries image-copyright and site-access/consent questions
(some heritage sites and museums prohibit commercial 3D capture; crowd-sourced tourist photos are
third-party copyrighted). The briefs treat the §17.1 *representational* care surface well but do not
add the legal-provenance gate on the capture inputs. For a ship-everywhere commercial title this
needs a per-site capture-rights check, parallel to the use_class discipline already in the
book-substrate lane. Name it; don't assume "real capture" = "cleared to ship."
strongest part of the corpus.
---
The owner ruled: (i) licensing never constrains shipping — operationalized as
unrestricted-components-by-default; (ii) Claude Code drives NeoStack schema-bound; (iii) the
game ships everywhere.
Tracing every shipped-asset path for a restricted component:
Indie (perpetual, commercial) → UE. CLEAN. (Caveat: Gaea CLI tier unconfirmed — cost, not a
ship-restriction.)
Audio Open (community <$1M). CLEAN / royalty-class thresholds only.
ones chosen.
nvdiffrec, which are NVIDIA non-commercial for third parties** (LOCAL_3D_ASSET_GEN.md §1.1 quotes
the license text directly; issue #22 unresolved, re-confirmed by Q3 §1.1). The ruling's escape is
"a clean texture leg picked at the benchmark." **Adversarial finding: that clean automated leg is
currently an EMPTY SET among confirmed open tools.** The candidates are all disqualified:
Hunyuan3D-2.1 paint = territory-capped (EU/UK/KR excluded), i.e. not unrestricted; Ubisoft CHORD
= Research-Only Copyleft, non-commercial (Q3 §1.2); Pixal3D = **inherits TRELLIS.2's nvdiffrast
chain (Q3 §1.2); the trellis-refactored PyTorch3D swap = archived/inactive since Nov 2025,
buggy. The only genuinely clean routes are manual** — Blender bake / ArmorLab (permissive but
human-authored). **So constraint (i) holds at the texture leg only by dropping automation, or by
accepting Hunyuan's territory cap.** The "AI textures the assets" efficiency assumption underneath
the pipeline is unfunded by any confirmed commercially-clean automated component. This is the
single most important ruled-constraint gap. It doesn't block shipping (you can always hand-author
PBR), but it silently converts an automated stage into a manual one — a schedule fact the corpus
doesn't foreground.
UE_BUILD_AUTOMATION.md §8.4 / headline #11: the UE EULA generative-AI clause restricts using the
Licensed Technology "as a training input [or prompt-based input] into any Generative AI Program."
The plain reading (which the brief endorses) is that this bars training on Epic content, not
using AI to author your content or drive the editor — so the NeoStack pipeline is fine. But the
"[or prompt-based input]" wording is genuinely ambiguous, and the pipeline's whole design **feeds
engine/asset context to Claude as prompt input.** The brief correctly triages this to a legal read
(ENGINEERING-SPIKES in RESEARCHED_STACK.md). **Verdict: not broken, but constraint (ii) rests on a
plain-reading of an ambiguous clause that has not had the legal read the brief itself calls for —
don't lock architecture as if it's cleared.** Separately, MetaHuman's own EULA hard-prohibits using
MetaHuman assets as AI training/test data (§8.3) — irrelevant to authoring, but a real fence if any
future pipeline stage ever fine-tunes on generated MetaHumans.
The operationalization is sound for tools: Hunyuan3D-2.1 and HY-World 2.0 (EU/UK/KR exclusions) and
Lyra 2.0 (non-commercial) are all correctly ruled previz-only, never load-bearing in the shipped path
— provided the corrections above (Lyra especially) are applied. The clean-default set (Copernicus,
MIT geometry, MetaHuman, Apache/MIT audio+SLM) genuinely ships worldwide. **The unaddressed
ship-everywhere threat is not a tool license — it's localization** (Docket 5): a worldwide narrative
game is not "shipped everywhere" in any meaningful sense without a localization pipeline, which no
brief researched.
Answer: the SDK code license does NOT cover the bundled model weights. I checked this session:
current sources describe the ACE Game Agent SDK code as MIT — which is itself a discrepancy
with Q3_2026_MODELS_REFRESH.md §3.5's claim of "Apache 2.0, directly quoted from the fetched
LICENSE.txt (VERIFIED)" (re-verify directly; one of the two is wrong). Either way, the ~3GB of bundled
weights carry their own separate licenses: Qwen3.5-4B (Apache 2.0), Chatterbox-Turbo-350M (MIT),
the NeMo-Conformer ASR model and the alternative Nemotron-3-Nano-4B (NVIDIA's own bespoke Open Model
License per Q3 §3.1, "read before shipping"). **Q3 §3.5's headline — "License: Apache 2.0… clears
the shipping-embedded bar cleanly" — conflates the SDK-code license with the weights license.** In
practice the default reference weights (Qwen/Chatterbox) are clean, so the conclusion isn't fatally
wrong, but the framing is the same error class as VIDEO_TO_3D's Lyra mistake — a code license read
as if it clears the weights. Two instances of the same slip in the corpus is a pattern worth a
standing check: for any AI model, read the weights license separately from the code license.
---
The acceptance chains and the game's own design pillars need these; no brief covers them:
1. Animation / motion generation. MetaHuman Animator covers human facial/body mocap-retarget,
but nothing covers **generative locomotion / motion synthesis for the 132-creature roster + 22
familiars + non-humanoid rigs** (quadrupeds, dragons, serpents). LOCAL_3D_ASSET_GEN.md §3.2 flags
rigging as unsolved and names paid point-solutions (Tripo universal-rig, AccuRIG, Sorceress) without
researching them. Motion/animation authoring at roster scale is a genuine unresearched lane and a
likely bottleneck.
2. Water / ocean / underwater. The endgame has drowned/submerged chapters (Lemuria/Atlantis-class)
plus coastal and river regions; region-template §1 carries a "water plane height" field. UE Water +
Landmass get only passing mentions (GEODATA §6, UE_BUILD §2.6). No lane on large-scale ocean,
buoyancy, underwater rendering, or the drowned-city aesthetic. Load-bearing gap.
3. Weather / sky / atmosphere / time-of-day. Region-template §18 is literally "Time Period /
Fluctuation." Earth-2 is climate-simulation (wrong domain, ruled out). Nothing covers UE Sky
Atmosphere / Volumetric Clouds / dynamic weather for 69 real-world regions across eras.
4. Crowd systems. MetaHuman Crowd is flagged Experimental with a "keep a classic modular fallback"
caveat (UE_BUILD §5.5), but Mass/MassEntity city-population crowds for the modern chapters are not
their own researched lane.
5. Localization / i18n — the biggest omission relative to constraint (iii). A ~30M-word (design
target ~90–120M) narrative game "shipping everywhere" has no localization lane at all: no
machine-translation-at-scale plan, no UE String Table / Localization Dashboard pipeline, and — most
importantly — no treatment of the **§17.1 cultural-care implications of translating culturally
sensitive content** into the languages of the very cultures being represented. This is arguably the
single largest gap for a ship-everywhere title.
6. Foliage / vegetation source assets. PCG Biome Core *scatters* vegetation, but the source
trees/plants (SpeedTree-class, or Megascans/Fab) are not a researched lane; generative 3D is
object-scale and weak on vegetation.
7. Chaos / destruction / physics authoring. "Vril-altered physics = real physics + real magic" is a
design pillar; no lane on Chaos destruction/fracture or the physics-authoring surface that pillar
implies.
8. VFX generation at scale. NeoStack can author Niagara, but there's no generative-VFX lane for the
per-region vril-signature effects across 69 regions.
(Minor, adjacent: no lane on the texture-streaming / disk-footprint / package-size budget for 69
real-world regions at AAA fidelity — a shipping constraint the throughput discussion never reaches.)
---
1. Fix RESEARCHED_STACK.md §6 and both briefs on Lyra 2.0: weights = NON-COMMERCIAL
research-only; Lyra is previz-only, never shipped-path. (VIDEO_TO_3D_BRIDGES §1.4.2 + §"genuinely
adds value" item 3.)
2. Strike "5.8 CONFIRMED SHIPPING" and "V2 stable, V3 announced" from RESEARCHED_STACK.md;
NeoStack-on-5.8 is UNCONFIRMED — keep the first-use spike as a real gate.
3. State the PBR-texture-leg hole explicitly in RESEARCHED_STACK.md: no confirmed clean automated
open text-to-PBR component exists; the clean routes are manual (Blender/ArmorLab) or territory-capped
(Hunyuan). Budget the texture stage as human-authored until the benchmark proves otherwise.
4. Correct NVIDIA_WORLDGEN's table (Cosmos 3 = 16B/64B, not 2B/14B) in place, not just in the sibling.
5. Re-verify the ACE SDK license directly (MIT vs Apache 2.0 conflict) and split code-license from
weights-license in Q3 §3.5; add a standing "read weights license separately" reviewer check.
6. Disambiguate "UE 5.8 MCP Server plugin" in MILESTONE_BUILD_PIPELINE.md (NeoStack-via-MCP-mode
≠ Epic's native 5.8 MCP plugin) before any autonomous loop parses it.
7. Cost the 69-region × hundreds-of-NPCs throughput (GDAL batch time, Gaea erosion time, MetaHuman
cloud rate-limit) — it is asserted, never modelled.
8. Open the missing lanes — localization first (ship-everywhere), then animation/motion, water,
weather; add a capture-rights gate to the reconstruction care rule.
---
*Adjudication complete. Corpus is source-grounded and disciplined; the actionable errors are
concentrated in (a) the Lyra license mis-read and (b) the consolidation layer running ahead of its own
briefs on NeoStack. Both are wired into the top-level docs Phase D would write against — fix those two
first.*