PIPELINE_READINESS_ROADMAP.md

pipelines/PIPELINE_READINESS_ROADMAP.md

Pipeline Readiness Roadmap

Status: ACTIVE v1.0 · 2026-07-19 · owner Josh · synthesized from the seven lane-audit reports in docs/pipeline_review/ (the full reports remain the workflow record; this doc is the consolidated map). CVD §17.14-conformant.

This doc answers Josh's question verbatim: "we have a lot of pipelines that we can test before the RTX 5090 computer comes in right, like music? what else isn't blocked? how are you building our combat and physics and other UE systems? I have NeoStack and Unreal already. when do you actually need the new computer and what's gated by book processing?"

The short answer

Yes — most pipelines are testable NOW on the current machine (UE 5.8.0 installed — BuildId 55116800, see §6; NeoStack installed as source, Python present). Six of the seven lanes have a real end-to-end smoke test that runs this week with zero new hardware. Music is one of them, though the marquee AIVA track is a manual-web-app path (no self-serve API), while the ElevenLabs SFX path is a clean automated loop.
Combat, physics, and the other runtime UE systems are built on the installed UE 5.8.0 engine using engine-native subsystems (Gameplay Ability System for Composure and the vril-density damage formula, StateTree for boss AI and the combat clocks, MetaSounds for the difficulty dial) plus a deterministic strategy-token runtime layer that is pure code and needs no GPU. The one true blocker here is a documentation gap, not a hardware gap: there is no combat/systems Translation Doc (lane: combat).
You actually need the 5090 for exactly four things: heavy local generative-model inference, at-scale batch bakes, performance-budget validation, and the benchmark days that pick defaults. Nothing else should ever be parked behind it.
Book processing gates content fidelity, never pipeline machinery. It only bites at un-scanned high-care regions and the pass-2 production quest/region content that depends on them; the Ch 2-5 slice and everything mechanical is already unblocked from the drained 246-book corpus.

1. NOW / 5090-GATED / BOOK-GATED dependency matrix

Every lane against its true gate. Evidence cited inline as (lane: name).

Lane	Unblocked NOW (current machine)	5090-gated	Book-gated	First-artifact true gate
Terrain	DEM download to manual UE Landscape import; needs a free OpenTopography key (Josh registers) + one rasterio install into Python 3.12 (lane: terrain)	Gaea erosion benchmark; 69-region throughput; World Partition + PCG perf-budget (lane: terrain)	None — geometry comes from Copernicus GLO-30 + WorldCover; machine layer is build-ready now (lane: terrain)	The missing machine-readable WGS84 region_bbox + single meters_per_vertex value (lane: terrain)
UE5 Build	Create the .uproject on UE 5.8.0; compile one USTRUCT; scripted DataTable ingest; PCG graph; World Partition scaffold; RemoteControl; NeoStack on 5.8 (source-built, verified 2026-07-19) (lane: ue5)	14B model for NeoStack; perf-budget across spec classes; 69-region headless bake; bulk terrain preprocess; MetaHuman batch staging (lane: ue5)	Per-region MetaHuman appearance authenticity + Section 16 provenance for un-scanned regions only (lane: ue5)	The per-registry USTRUCT column-to-UE-type spec + the spine-row extractor tool (lane: ue5)
Asset generation	Blender-headless finishing chain; FBX-to-UE import + registry writeback; NeoStack previz passthrough; single MetaHuman; Fab acquisition (lane: assetgen)	All local geometry generation (TRELLIS.2/Hi3DGen/Hunyuan3D, 24GB+ floor); the benchmark protocol; the PBR renderer-swap spike; bulk volume (lane: assetgen)	Low — reads region pages not raw books; soft ties are capture-rights photos + creature-roster completeness trailing the drain (lane: assetgen)	The mesh_tags object + per-asset routing tag + capture-rights field, none present as schema (lane: assetgen)
Audio (music/voice/SFX)	ElevenLabs SFX to UE import (automated); Kokoro local voice; MetaSound Builder spike (lane: audio)	ACE-Step/YuE bulk music; instrumental benchmark; MOSS-SoundEffect/Stable Audio batch; Demucs; Kokoro bulk throughput; adaptive-graph perf (lane: audio)	Verification-grade only, at high-care cultures (Sámi joik Ch 29, Vodou Ch 42, songline Ch 57); NOT generation-grade (lane: audio)	Dialogue line text unauthored (from Quest §6) (lane: audio) — the AIVA API wall is moot: closed as skip 2026-07-27
Cinematic	Template Level Sequence + headless driver; MRQ 720p render; single MetaHuman; DataTable ingest; dispatch stub (lane: cinematic)	Render fidelity (Lumen/Nanite/groom LOD0); many-MetaHuman perf; batch dailies; roster-scale MetaHuman (lane: cinematic)	Not meaningfully — only the batch cultural audit, covered by the drained corpus (lane: cinematic)	scene_type-to-shot-grammar mapping absent; the real cross-lane wall is audio (lane: cinematic)
Quest writing	Author Ch 2 quest rows; Ch 3-5 SPIKE; a sample dialogue_dataset; critic pass; baseline refresh (lane: quest)	Nothing for authoring; production volume helps via draft-tier pre-draft + bulk FK + embedding dedupe (lane: quest)	Ch 2-5 SPIKE is NOT book-gated; pass-2 per-region production authoring is region-page-gated, so book-gated at high-care regions (lane: quest)	Three unpinned decisions: dialogue-dataset format, landing path, condition_expression grammar (lane: quest)
Combat / physics / runtime	GAS greybox boss; StateTree clocks + habit-vector; vril_density dial; deterministic strategy-token spine; cosmetic barks on the current GPU (lane: combat)	14B reference model; runtime-inference perf-budget; heavy asset-gen for vertical slice; nightly eval at scale (lane: combat)	Fauna behavior_flavor partial-now; §17.1 runtime denylist accretes; per-region encounter population is region-page-gated (lane: combat)	No combat/systems Translation Doc and no combat section in the contract (lane: combat)

2. The unblocked-now execution queue

Ordered by leverage then cost. The already-ruled items are slotted in place: the Ch 2-5 region-page slice (pass-1 plus deep-research fill, ruled 2026-07-19 as Option A plus the slice exception), the quest-writing SPIKE (validation artifact only — production quest content per region is book-gated pass-2), the RCM substrate order, RR-62, and the D-1 consistency audit.

#	Step	Machine	Prerequisites to install / configure / key	Doc it validates	Exit criterion
1	Quest-writing Ch 2 SPIKE — one main-story quest row appended to T0_Quest_Definition_Registry plus one sample dialogue_dataset (six-option node), critic-gated. Validation artifact only (lane: quest)	Current — Opus authoring pass, no hardware	Nothing beyond the repo and this session	T1_Build_Pipeline_Contracts §6	Row conforms, run_gates.py green (unpiped) after baseline refresh, normalize-doc GO, and the three quest WALLs forced to a decision
2	RR-62 range review — reveal-discipline plus cultural-care fan-out across the range, fixes applied (parallel canon verification)	Current — Opus fan-out, no hardware	normalize-doc skill; read-only critics on the shared tree	Spine reveal-discipline + Care Doctrine (CVD §17.1)	RR-62 GO with all found leaks fixed and re-verified
3	D-1 consistency audit — resolve or queue the D-1 seating/consistency items across the affected nodes (parallel)	Current — Opus, no hardware	The D-NTALA-ENC / D-1 ruling record	The D-1 ruling and adjacent spine/registry consistency	Audit complete; inconsistencies resolved in-place or queued to DECISIONS_PENDING_JOSH
4	RCM substrate order — issue the deep-research substrate pulls for the Ch 2-5 slice (default synthesis path; MUST-ACQUIRE physical buys reserved to the curated set)	Current — deep-research agents	Confirm the two Flores MUST-ACQUIRE titles (Erb; Allerton) against BOOK_ACQUISITION.md	T1_Research_Corpus_Master + docs/BOOK_ACQUISITION.md posture	Substrate packs available for the four region pages; MUST-ACQUIRE confirmed, everything else synthesized
5	Ch 2-5 region-page slice — pass-1 build plus deep-research fill (ruled Option A + slice exception, 2026-07-19)	Current — Opus + deep-research	The slice substrate from #4; region pages bali and sumatra_java already present	T1_Region_Page_Template v2.0	angkor_khmer built (page 4) + all four slice pages deep-research-filled to review depth (flores/bali/sumatra_java pass-1 already landed); Section 5/6 high-care depth pending the MUST-ACQUIRE books
6	Combat continuo-bet runtime smoke — llama.cpp 3-4B GGUF under a GBNF grammar, through the deterministic V1-V5 validator plus authored fallback plus pass-rate log (lane: combat)	Current GPU (3-4B class)	llama.cpp build + one GGUF download (Qwen3-3B / Phi-4-mini class); hand-author one boss context pack	RUNTIME_GUARDRAIL_ARCHITECTURE.md + RUNTIME_GENERATIVE_LAYER.md	The loop never emits an invalid or off-whitelist line (fallback always catches) and prints a first-try pass rate
7	UE5 registry-to-DataTable-to-PCG-to-rendered-zone smoke — compile FCreatureRow, scripted CSV import, PCG scatter, MRQ frame (lane: ue5)	Current — UE 5.8.0	Confirm VS 2026 "Game development with C++" workload once before first compile	UE_BUILD_AUTOMATION.md + T99_Translation_UE5_Build	Cubes scatter per creature row in a World Partition level; screenshot or MRQ frame captured
8	Audio ElevenLabs SFX-to-UE import round-trip — Flores Section-10-grounded 48kHz WAV imported via SoundFactory (lane: audio)	Current — UE 5.8.0	ElevenLabs API key + credits (short clip fits the free tier; full 30s needs a paid tier)	T99_Translation_Audio + Contracts §10	Canon-substrate-grounded WAV living at /Game/Audio/SFX/environmental_ambient/
9	Terrain download-preprocess-manual-import — Liang Bua COP30 to a UE Landscape (lane: terrain)	Current — Python + UE 5.8.0	Josh registers a free OpenTopography key (account creation is his to do); pip install rasterio into Python 3.12	GEODATA_TERRAIN_CHAIN.md + T99_Translation_Terrain	Liang Bua relief visible in a UE Landscape; surfaces the true bbox and meters-per-vertex numbers the page lacks
10	Asset-gen finishing plus FBX-to-UE import plus registry writeback — one previz fantastical prop (lane: assetgen)	Current — UE + Blender	Blender installed; NeoStack cloud allowance; a throwaway input mesh	T99_Translation_AssetGen §4.3	mesh_id_ref written back on the owning row; the missing §2.6 schema fields enumerated
11	Cinematic template Level Sequence plus headless driver plus MRQ render — one scene row to a 720p video with placeholders (lane: cinematic)	Current — UE	One hand-authored scene_type template sequence	T99_Translation_Cinematic	A rendered video with placeholders on the templated camera
12	Combat GAS greybox difficulty-as-density proof — one Composure AttributeSet and a single vril_density float that widens/narrows the tell window with no enemy-stat change (lane: combat)	Current — UE 5.8.0 (GAS engine-native; the 5.8 AttributeSet-init deprecation now applies — verify the path)	Author against GAS AttributeSets/GameplayEffects; avoid the 5.8-deprecated attribute-set-init functions	COMBAT_ENCOUNTER_SYSTEM.md §5/§9 + the RULED difficulty system	The same authored boss plays at two world-tensions with zero change to any enemy stat block

Notes on the ruled items: the quest SPIKE (#1) is deliberately a validation artifact — it proves the lane end-to-end and forces the three format/path/grammar decisions; per-region production quest content is pass-2 and book-gated where its region page is book-gated (lane: quest). The region-page slice (#5) builds the machine layer now and defers Section 5/6 cultural depth to the two MUST-ACQUIRE Flores books (lane: terrain, lane: quest). Items 1-5, 11 and 12 need no purchases; items 6-10 need one small install or one free/low-cost key each.

3. The wall list

Every parameter gap rated WALL across the lanes, its owning doc, and the fix route. The missing combat-systems Translation Doc is confirmed by the combat lane report and called out first because it blocks the entire combat/physics/runtime engine-mapping decision.

WALL	Lane	Owning doc	Fix route
✅ RESOLVED (2026-07-19): T99_Translation_Combat.md [DRAFT v0.1] LANDED — authored to the five-doc bar, 3-critic adversarially gated (14 fixes applied incl. the BYPASS win-verb restoration, the boss_kind-not-boss_scale reconcile, the WS_024 [0,2]→GAS normalization, seam tags, the §8.1 PARAMETERS BLOCK). The GAS/StateTree/Chaos forks are resolved IN the doc; its 14 [JOSH] canon flags ride the queue as D-COMBAT-TD; the contract combat-§ pointer + the Boss_Encounter column extension apply after those rule	combat	docs/translation/T99_Translation_Combat.md	The former wall; remaining follow-ons are D-COMBAT-TD rulings + the greybox prototype (queue #12)
T0_Boss_Encounter_Registry lacks the machine combat-spec columns — no primary/secondary verb, Patterns/Puzzles/Preparation axes, win-condition enum, per-signature counters, Composure nodes, or win-and-death reward fields (the spine §9 thin-index carries verbs/axis as prose rows, not machine columns). NOTE: the registry IS populated — 281 rows — the lane report's 0-rows claim was refuted at director adjudication; the wall is schema fitness, not emptiness	combat	T0_Boss_Encounter_Registry + COMBAT_ENCOUNTER_SYSTEM.md Gap 3.2	New spec doc + a registry schema EXTENSION (declared via docs/registry_extensions.json, fidelity re-baselined) aligning the 281 rows + the boss_roster_rows.json proposal to machine combat columns
Every combat number is deferred to a Phase-5M balance audit that never ran — damage constants, integrity-band absolute units, sigma values, tell/window ms, dial coefficients	combat	COMBAT_ENCOUNTER_SYSTEM.md	Run the balance audit as its own workflow; greybox proceeds on illustrative shapes, a tuned build cannot
No canon win-condition taxonomy, no mechanical definition of the mandatory character-engagement field, no source authority for the three-axis split	combat	COMBAT_ENCOUNTER_SYSTEM.md (Pillar-8)	Josh canon ruling + spec doc defining the axes and win-condition enum
The recall-PACKS offline exporter (the frozen per-boss context-pack builder the guardrail architecture depends on) does not exist. NOTE: harness/recall.py itself EXISTS and works — the lane report's claim it is absent was refuted at director adjudication; the missing piece is the pack-export layer on top of it	combat	RUNTIME_GENERATIVE_LAYER.md §C + RUNTIME_GUARDRAIL_ARCHITECTURE.md §9	Build the packs exporter in harness/ atop recall.py; a prototype hand-authors one pack, production population cannot start until it exists
No per-registry USTRUCT spec maps each CSV column to a UE C++ type (FText/FString/int32/float, enums, FDataTableRowHandle FKs, TArray/JSON for nested cols)	ue5	T1_Build_Pipeline_Contracts §1.5-1.6	Contract § or new spec doc — add the column-to-UE-type contract the codegen needs
The spine-row extractor (markdown-table to CSV pulling site/beat/fork/scene/SQ rows from docs/spine/CH_NN.md) does not exist; enrich.py and extract_claims.py do not produce it	ue5	harness/ (absent tool)	Build the extractor in harness/ — it is the §3 pre-step no UE step can run without
PCG population numbers (ambient_threat-to-density, vril_density-to-tier filter, climactic_affordances-to-markers) are numeric inference assigned to the Translation Doc layer, which is unpopulated; no terrain_type-to-PCG-template binding table exists	ue5	T1_Build_Pipeline_Contracts §0 (Translation Doc layer)	Populate the inference layer as Flores surfaces cases; author the terrain_type-to-PCG-template binding table
The 11 scene_type shot-template Level Sequences do not exist; camera blocking/timing/cut rhythm is not in the seven-column schema; no dialogue/audio (SoundWave FK) rides the scene row	ue5, cinematic	T99_Translation_UE5_Build §3 Step 9 + T99_Translation_Cinematic + SPINE_SCHEMA_v3	Author the 11 shot templates (hand or NeoStack) + add camera-block and SoundWave-FK fields to the scene-row schema
The mesh_tags object (morphology/region/style/material/era) exists in no consumable form — not a structured block on the region page, not populated on live registries	assetgen	T1_Build_Pipeline_Contracts §2.5 + T1_Region_Page_Template	Template field add + registry columns — instantiate mesh_tags per asset (it is the primary prompt-composition input)
No machine-readable per-asset asset-class / mesh_category routing tag — the cloud/local/reference-built split lives only in Section prose	assetgen	T1_Region_Page_Template + Contracts §2	Per-row field — add mesh_category so backend routing is deterministic
No reference_image_ref / capture-rights / use_class field for the §17.1 reference-built architecture path; the agent cannot distinguish a rights-cleared photo from a prohibited one, or from AI-generated input	assetgen	T1_Build_Pipeline_Contracts §2.11/§6 + region page	Template field + contract § — add the capture-rights-cleared source FK before any real cultural architecture is reference-built
~~No self-serve AIVA API exists~~ — CLOSED AS SKIP 2026-07-27, no longer a gap: the vendor lane is dead	audio	T99_Translation_Audio Music Step 2 + Contracts §3	Ruled, not fixed — the canonical and hero theme track is composer-in-loop with a human composer; bulk music is local ACE-Step/YuE, benchmark-gated
No T0_Theme_Registry field carries tempo, bar-length, or stem-role metadata — MetaSounds+Quartz cannot consume a cue as a true adaptive layer	audio	T99_Translation_Audio Music Step 4 + T0_Theme_Registry	Registry field add — ratify the PROPOSED tempo/bar/stem-role fields
Dialogue line text does not exist anywhere — dialogue_dataset_ref is populated by the unauthored Quest Writing pipeline, so no canonical voice line can be generated today	audio	Quest Writing §6 + dialogue_dataset	Cross-lane — author quest content (queue #1 starts this); Voice Step 1 derives who speaks, never what
The dialogue_dataset file format and extension are unpinned by the contract	quest	T1_Build_Pipeline_Contracts §6.4 + PIPELINE_ADAPTER.md §6	Josh pin — decide format (JSON natural) and extension together at the first quest authored
The dialogue_dataset landing path is a dead Google-Drive path; the repo replacement canon/dialogue_datasets/ is an un-ratified proposal and the dir does not exist	quest	PIPELINE_ADAPTER.md line 120	Ratify the landing path + create the directory
The condition_expression grammar (runtime branch conditions) is undocumented — operands are named but no expression syntax exists	quest	T1_Build_Pipeline_Contracts §6.11	New grammar spec or contract § defining the condition_expression syntax
No machine-readable WGS84 region_bbox — the golden page carries free-text local-ENU design sizings, and the real bbox is explicitly deferred to the Section 1 terrain pass	terrain	flores_island.md + T99_Translation_Terrain Step 1 / PARAMETERS BLOCK	Template field add — put a WGS84 south/north/west/east (or center+radius) per zone into Section 1 / the Zone Catalog
No instantiated alteration_list — the §3 carve/composite schema is specified but no fillable recipe exists on any golden page (operation enum, feather_radius_m, WGS84 footprint all absent)	terrain	T99_Translation_Terrain §3	Instantiate the §3 schema on the region page (depends on the bbox fix above)
No single resolved meters_per_vertex value — the rule is one number per region but the page gives multi-valued prose	terrain	T99_Translation_Terrain §5 item 1 (JOSH-RULES)	Josh ruling + template field for the one machine value
participant CHAR_id has no MetaHuman asset-path lookup — depends on the UE5 build having run Step 7 and emitted stable IDs; no UE build has run	cinematic	UE5_Build Step 7 dependency	Cross-lane — the UE build must generate MetaHumans and emit stable asset IDs first
convergence_node_scene_dispatch schema is unratified and the ws.* WorldStateSubsystem does not exist — every LOCKED/path-conditional scene cannot clear	cinematic	T99_Translation_Cinematic spikes 3/7 + the (absent) WorldStateSubsystem	Ratify the dispatch schema + build the ws.* subsystem (shared with combat save state)
Boss cinematics fire off a boss combat-state event that no boss-AI framework yet emits	cinematic	T99_Translation_Cinematic spike 4	Cross-lane — depends on the combat lane's boss-AI framework (StateTree) emitting the state event
Ambient scene rows carry no beat_ref, so the beat-to-site-to-region join has nothing to join from	cinematic	T99_Translation_Cinematic spike 6	Schema field add — put a beat_ref (or a specified substitute) on ambient rows
Body-performance needs a mocap/retarget library that does not exist, and bespoke non-audio hero-beat animation has no pipeline	cinematic	T99_Translation_Cinematic spikes 9/2	Build or acquire a mocap/retarget library; define the hero-beat animation pipeline

4. What the 5090 is actually for

Consolidated and deduped across all seven lanes. If a task is not on this list, it does not wait for the workstation.

Heavy local generative-model inference. Local geometry generation clears a 24GB VRAM floor (TRELLIS.2 needs 24GB, Hunyuan3D-2.1 up to 29GB) that the current GPU cannot meet (lane: assetgen). Bulk instrumental music needs the same class (YuE ~24GB, ACE-Step XL 12-20GB) (lane: audio). The 14B on-device reference model that the NeoStack interpretive seam and the runtime generative layer assume for dev-quality prototyping is reserved to the 5090 by the design's own pin (lane: ue5, lane: combat); the 3-4B ship-target barks model already runs on the current GPU.
At-scale batch bakes. The 69-region headless PCGWorldPartitionBuilder plus Nanite/HLOD at 8129-vertex landscapes is memory- and time-bound (lane: ue5). Bulk terrain preprocess (GDAL to Gaea erosion synthesis to PNG16) and the kilometre-scale Gaea erosion sims are heavy local compute (lane: terrain, lane: ue5). Bulk asset-gen at real volume (72 weapon base meshes, 132 creatures, 33 familiar species, vril anchors) and MetaHuman generation at roster scale live here (lane: assetgen, lane: cinematic). Bulk SFX (MOSS-SoundEffect, Stable Audio Open) and Kokoro voice at full per-chapter/per-region scale are the batch tier (lane: audio).
Performance-budget validation. Runtime-inference budgets — batching, speculative/draft decode, KV/prefix caching, the ~8-15 gen/s target, latency P99s (lane: combat). UE render and streaming budgets across the four spec classes (steam_deck_minimum/mid/high/ultra), triangle/draw-call/streaming, Lumen/Nanite/strand-groom at LOD0, many-simultaneous-MetaHuman groom cost, and the MetaSounds/Quartz adaptive-graph memory and decode cost (lane: ue5, lane: cinematic, lane: audio).
Benchmark days that pick defaults. Gaea-vs-UE-Landmass erosion quality (decides whether Gaea stays in the chain) (lane: terrain). The asset-gen §5 protocol on the actual 5090 (per-category default + time-to-game-ready) plus the PyTorch3D PBR renderer-swap spike (lane: assetgen). The instrumental-orchestral music model bench (AIVA vs ACE-Step vs YuE vs HeartMuLa) and the Demucs stem-separation fitness test (lane: audio). The nightly first-try guardrail-pass-rate eval at corpus scale (lane: combat).

Local-model fine-tuning and distillation (ADDED 2026-07-19, reconciled from SELF_IMPROVING_QUALITY_LOOPS §4.3 — the one genuine model-training workload): LoRA/QLoRA on a 7-14B local base over our own ratified corpus (first target: a persona-voice model mapping persona cards to in-voice lines) and distillation of the 14B reference into the 3-4B ship model. Scale-bounded by the 32GB card — no full large-model fine-tune. Dataset accumulation starts NOW on the current machine: ratified personas + gate-passed runtime lines as positives, the fallback bank and validator-rejected lines as hard negatives — every training row ratified or gate-passed, never model-generated canon (lane: combat, lane: quest).

Console packaging (Xbox int4/DirectML, PS5 compute-shader) is separately devkit-gated — neither the current machine nor the 5090 (lane: combat).

Keys & external services — the ruled shortlist (2026-07-19)

The complete set of external keys/subscriptions worth acquiring now, and the explicit skips. Nothing outside this list blocks the §2 NOW queue.

OpenTopography — ✅ KEY IN + PROBE-VERIFIED (2026-07-20): a real Liang Bua COP30 GeoTIFF pulled through it, and the queue-#9 download+preprocess legs are now SCRIPTED AND RUN (game repo Tools/terrain_fetch_preprocess.py: fetch → UTM reproject at chosen m/vertex → 16-bit PNG heightmap + the machine-readable meta JSON — Liang Bua: bbox −8.58..−8.48 / 120.39..120.49, EPSG:32751, 30 m/vertex, 371×372 px, elevation 141.75–1187.07 m, UE Z-scale 204.164 — exactly the region_bbox + meters_per_vertex numbers the §3 wall named). The UE Landscape import leg rides the next editor session. Mind the 50-calls/24h cap; the keyless Copernicus AWS fallback stands for bulk.
ElevenLabs — ✅ KEY IN + PROBE-VERIFIED (2026-07-20): Starter plan, 90k credits; the SFX generate leg is SCRIPTED AND RUN (game repo Tools/generate_sfx.py, first artifact smoke_stone_door.mp3). The →UE import leg rides the next editor session. Final VO and bulk SFX remain the 5090/local tier.
Key-storage note: both keys live in the user registry (HKCU:\Environment, via setx). Shells whose parent process predates the setx do NOT inherit them — the Tools scripts read the env var first and FALL BACK to the registry, so they work from any session.
AIVA — CLOSED AS SKIP, RULED by Josh 2026-07-27 (“not emailing; the enterprise-contact path is dead”); the 2026-07-19 recommendation that follows is what he ruled on. Three reasons: no self-serve API exists (§3 wall — the path is manual-web-app, so it automates nothing); the intended music default is the open-source bench class (ACE-Step / YuE / MusicGen / Stable Audio Open), which is a 5090 benchmark-day decision, local and license-clean with no subscription; and the canonical-theme track is composer-in-loop regardless. AIVA does not re-enter — Josh closed it as a skip on 2026-07-27. If the open-source bench disappoints, the fallback is wider composer-in-loop coverage, not a vendor return.
Nothing else. Cloud asset-generation choices (Meshy/Tripo passthrough) ride the assetgen benchmark, not a key purchase today; MetaHuman Creator Core Data is a later in-editor install, not a key.

5. What book processing actually gates

Book processing gates content fidelity, never pipeline machinery. Nothing deterministic in any lane waits on the drain. What it does gate:

Region-page Build-Ready dates for un-scanned high-care regions only — Ch 29 Sámi, Ch 42 Haiti/Vodou, Ch 57 Australia, and Ch 11, 12, 13, 47-48, 50 — which depend on the roughly three remaining weeks of Josh's scanning. Most non-high-care sources are deep-research-synthesized per BOOK_ACQUISITION.md, not book-gated (lane: quest, lane: ue5).
Pass-2 production quest authoring, which triggers on region-page Build-Ready: books gate the region pages, and the region pages gate the regional quest set. The main-story and hard-line quest set is spine-gated, not book-gated (lane: quest).
MetaHuman appearance-range authenticity and Section 16 citation provenance for un-scanned regions — the fidelity of what a built region depicts, not the deterministic UE spine (lane: ue5).
Verification-grade audio cultural-authenticity substrate at the high-care cultures (joik, Vodou drum/chant, songline). This is verification-grade, not generation-grade — a first Flores artifact needs no new books (lane: audio).
The §17.1 runtime cultural denylist / forbidden-framing corpus that the runtime cultural-safety gate matches against — it accretes as books and region pages land, so a combat or fauna build in a high-care chapter should not run runtime generation until that entity's care-band denylist is populated (lane: combat).
Fauna behavior_flavor and folklore rosters partially trail the drain — the 246 owned books are already drained and queryable, so this is partial-now; new folklore-creature and historical-figure entries queue as pending as scanning continues (lane: combat, lane: assetgen).

Not book-gated at all: terrain (geometry is DEM-derived) (lane: terrain); the cinematic deterministic spine (lane: cinematic); anything in the Indonesia arc (flores, bali, sumatra_java pages exist and the corpus is drained) (lane: ue5, lane: quest); the Ch 2-5 slice (lane: quest).

6. Stale-doc corrections

Every stale-or-wrong claim surfaced by the lanes, with its file, for living-source fix-up. The single biggest one — the UE 5.8-vs-5.7 split — recurs across four lanes and is listed once.

Engine-version split (recurs across lanes)

The terrain chain, T99_Translation_Terrain, T1_Region_Page_Template Section 1, T99_Translation_UE5_Build §2.1, T1_Region_Page_Template §17 title, T99_Translation_Cinematic, and T99_Translation_AssetGen all target UE 5.8, but T1_Build_Pipeline_Contracts §1 pins UE 5.7 and the only engine installed is UE 5.7. UE 5.8 is real and shipped, but nothing on the current machine runs the 5.8-only parts (Epic first-party MCP + PCGToolset + the PCG graph-generation skill, Mesh Terrain, Movie Render Graph "Production-Ready", MetaHuman 5.8 face coefficients). Fix: flag the split in each doc — the deterministic spine is fully 5.7-capable; the 5.8-only mechanisms need an engine install. Files: GEODATA_TERRAIN_CHAIN.md, T99_Translation_Terrain.md, T99_Translation_UE5_Build.md, T99_Translation_Cinematic.md, T99_Translation_AssetGen.md, T1_Region_Page_Template [ACTIVE v2.0].md, T1_Build_Pipeline_Contracts [ACTIVE v1.0].md.
On the installed 5.7 the NeoStack seam is more viable than the 5.8-pinned plan implies: 5.7 matches both the contract's stated NeoStack version and NeoStack's own confirmed 5.5-5.7 support range, so the compat spike exists only because of the 5.8 pin. Fix: note that on 5.7 the NeoStack seam sidesteps the #1 gating spike. Files: T1_Build_Pipeline_Contracts §1.10/§1.14, T99_Translation_UE5_Build.md.
SUPERSEDED SAME-DAY (2026-07-19, later): Josh began the UE 5.8 upgrade AND the vendor page (betide.studio Agent Integration Kit) ships an OFFICIAL pre-built NeoStackAI-UE5.8-v2.0.42 build (Jun 4 2026 builds for 5.5/5.6/5.7/5.8; Josh downloaded the 5.8 build directly). The NeoStack-on-5.8 spike (STACK_FACTS Q1) therefore collapses from "unverified support" to a ten-minute in-install verification: plugin loads clean in a 5.8 project + one trivial Blueprint-generation prompt round-trips. UE 5.7 stays installed side-by-side only until that verification passes, then the 5.8 pin becomes true across the docs and the engine-split corrections above apply in 5.8's favor.
RESOLVED — VERIFIED 2026-07-19 (final; supersedes the two bullets above): UE 5.8.0 (BuildId 55116800) is now the installed and pinned engine (project C:\dev\Humanity\Humanity, repo github.com/milkman3790/Humanity, branch master, LFS wired); the 5.7 side-by-side is retired and the engine-split corrections above now apply in 5.8's favor across every doc. The NeoStack-on-5.8 spike is PASSED — and the prebuilt path in the bullet above is corrected: NeoStack is installed AS SOURCE in the project Plugins/ and rebuilt via Engine\Build\BatchFiles\Build.bat against that exact BuildId — a prebuilt or marketplace binary drop-in re-breaks on a BuildId mismatch, so prebuilt drops are forbidden (the Jun-4 vendor zip is not the install vector). The NeoStack MCP channel is VERIFIED LIVE (http://127.0.0.1:9315/mcp, streamable HTTP, server unreal-editor, one execute_script Lua meta-tool over ~35 domains; the MCP-Session-Id header is required after initialize) and smoke-tested end-to-end (actor spawn + viewport screenshot). Epic's ModelContextProtocol plugin is enabled (server unreal-mcp, http://127.0.0.1:8000/mcp) but NOT auto-started and no toolset is enabled yet — still pending. NeoStack Agent Chat / Claude ACP is root-caused (the sign-in terminal stall is a known NeoStack auth-parser bug; working practice = ignore the sign-in terminal and just send a message; CLAUDE_CODE_EXECUTABLE is pinned in user env). The three-seam routing (NeoStack judgment seam · UE-native Python + Remote Control bulk seam · Epic PCG seam) is confirmed as the operating shape. Authoritative record: C:\dev\Humanity\Humanity\CLAUDE.md, "Editor automation channels (verified live, 2026-07-19)".

Terrain lane

GEODATA_TERRAIN_CHAIN.md dates UE 5.8 as shipped July 2026; actual GA was June 17, 2026. Minor date error; 5.8 existence is correct. File: GEODATA_TERRAIN_CHAIN.md §3/intro.
OpenTopography rate-limit and commercial-key terms are under-specified — the docs say only "free, issued instantly," omitting the 50-calls/24h non-academic cap and the Enterprise-key requirement for commercial software integration. Fix: record both. Files: GEODATA_TERRAIN_CHAIN.md, T99_Translation_Terrain.md.
Confirmed current, no action: Gaea 3.0 WATCH flag and the Indie CLI-tier caveat hold; PCG Biome Core "Experimental v0.2" is accurate across 5.5-5.8.

UE5 Build lane

UE_BUILD_AUTOMATION.md §1.1 claims embedded Python 3.11.8 is a 5.8 jump from 3.9.7 and says to re-verify older scripts — false on this machine: UE 5.7 already ships Python 3.11.8. Fix: drop the 5.8 version-gating on the Python bump. File: UE_BUILD_AUTOMATION.md §1.1.
NEOSTACK_AI.md §8 states Epic's native MCP ships no PCG toolset — stale on the PCG point: a dedicated PCGToolset plugin and a mandatory PCG graph-generation skill do ship for the 5.8 MCP. The base out-of-box toolsets are correctly narrow (Scene/Actor/Material/Object). Fix: correct the PCG half; UE_BUILD_AUTOMATION §3.6 and the translation doc §2.3 are the correct account. File: NEOSTACK_AI.md §8.

Asset generation lane

T1_Region_Page_Template Section 2 and the golden pages treat Fab/Megascans as a cheap/free reference source — stale: Epic began charging for Megascans in 2025, only ~1,500 assets remain free. Still commercially usable under the Fab Standard License, so a cost axis, not a licensing wall. Files: T1_Region_Page_Template Section 2, flores_island.md S2.
T99_Translation_AssetGen §3 frames the PBR-texture hole as an empty set — that predates the current TRELLIS.2, which now advertises CuMesh UV unwrapping plus full PBR materials (MIT). The nvdiffrast/nvdiffrec non-commercial caveat still holds, so the hole is not proven closed, but the §5 pre-flight recheck is now materially more urgent. Fix: re-verify against current TRELLIS.2 before repeating "empty set." File: T99_Translation_AssetGen.md §3.
T99_Translation_AssetGen §2 license tail treats Tripo's shipped-output terms as unverified — now resolved: Tripo Pro grants full commercial IP ownership of generated and auto-rigged assets, usable in Unreal (Free tier is non-commercial). Fix: close the settle-item at Pro tier. File: T99_Translation_AssetGen.md §2.
Confirmed current, no action: MetaHuman sunset dates, in-engine-native-since-5.6, and Hunyuan3D-2.1's MAU/territory cap all hold.

Audio lane

T99_Translation_Audio §2.6 and §4 spike #7 flag Kokoro's language coverage as unconfirmed — now confirmed: 54 voices across 8 languages at 24kHz. A real shortfall against a ship-everywhere mandate, adequate for a first artifact. Fix: update the spike from "unconfirmed" to "confirmed 8-language ceiling." File: T99_Translation_Audio.md §2.6/§4.
Moot after the 2026-07-27 AIVA skip ruling: the per-tier download cap versus composer duration-control nuance no longer applies to any live lane. File: T99_Translation_Audio.md.
Otherwise nothing in the audio docs is contradicted by current sources (live-re-verified 2026-07-19).

Cinematic lane

T99_Translation_Cinematic §2 treats batch audio-driven MetaHuman facial as fully solved — 5.8 ships a real-time model but does not declare batch-API maturity, and Epic's roadmap still tags audio-driven animation Experimental. Fix: add a verify-before-roster-scale caveat. File: T99_Translation_Cinematic.md §2.
T99_Translation_Cinematic §2.6 says T0_Theme_Registry and T0_Voice_Registry are not yet created — they now exist as DRAFT v0.1 stubs (Voice 11 lines, Theme 13), SFX ACTIVE v1.0. Correct in spirit (unpopulated), stale in wording. File: T99_Translation_Cinematic.md §2.6.

Quest writing lane

The dialogue-dataset Drive path /04-Quest Writing/Dialogue Datasets/ is dead (canon migrated off Drive); the canon/dialogue_datasets/ remap is un-ratified. File: T1_Build_Pipeline_Contracts §6.4/§6.11.
§6.3/§6.4/§6.6 read "region page Section 14 NPC dialogue substrate," but template v2.0 Section 14 is "Trade Skills, Detailed Content" and the cast lives in Section 12 — the pointed-at substrate does not exist as written. Fix: repoint quest authoring to Section 12. File: T1_Build_Pipeline_Contracts §6.3/§6.4/§6.6.
§6.4.1 and the status machine say "author reviews at the Drive doc surface" — no Drive exists; reviews are repo branch diffs. File: T1_Build_Pipeline_Contracts §6.4.1.
§6.2 still carries stale ending vocabulary ("Neutral/Bad ending"); the live registry is already reconciled to TRULY_GOOD/GOOD_ENOUGH/EVIL/UNLOCKED/ALL. Fix: grep-fix the prose to match the data. File: T1_Build_Pipeline_Contracts §6.2.
Nuance, not stale: the contract phrase "ElevenLabs v3 emotion control schema" implies a formalism that does not exist — v3 audio_tags are free-form bracketed strings (v3 is alpha/research-preview), not a fixed enum. Fix: reword to "free-form audio tags." File: T1_Build_Pipeline_Contracts (quest §).

Wiring

This doc is the consolidated readiness map; the seven full lane reports are landed verbatim at docs/pipeline_review/READINESS_AUDIT_2026-07-19_LANE_REPORTS.json (the uncorrected workflow record — two combat-lane claims, Boss_Encounter-0-rows and recall.py-absent, were refuted at director adjudication against the live tree and corrected in Sections 1-3 above). Add this file to docs/DOC_MAP.md (a doc is not done until it is listed there). The stale-doc corrections in Section 6 are living-source fix-ups — apply them in place per the living-source doctrine; any that touch chapter/numerical/naming/hard-line/§17 canon go to Josh, but every item above is a factual/currency correction, not a canon call. The WALL fixes in Section 3 that add template fields, contract sections, or new spec docs are the ordered unblock work behind the execution queue.

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