PIPELINE_REVIEW_MISSION.md

pipelines/PIPELINE_REVIEW_MISSION.md

The AI-pipeline review — the Fable-5 mission after the spine [ACTIVE, 2026-07-15]

Josh's direction (2026-07-15, verbatim intent): the Fable-5 job now is to get the AI pipeline

processing REVIEWED AND CORRECTED so Fable knows exactly how the tech stack works and how the AI

pipelines will consume the region pages — and therefore everything the region pages need. The

region-page templates, book processing, and downstream build pipelines exist but are NOT flushed

out, never had a Fable pass, and were written before (and are unaware of) the completed v3 story

spine. Parts of the stack written into those pipelines are being RETIRED.

The end state the pipelines must reach

From the RCM + the Story Spine (v3.1) + the Region Pages, the AI pipelines move AUTONOMOUSLY,

complete and end-to-end:

1. GIS / 3D world generation AND editing;

2. into Unreal Engine to build the world;

3. NPCs, quests, cinematics — the pipelines know how to plug in the quests, the characters, and

the cinema (the v3 event graph is the machine contract: beats→forks→state, §3b sites/terrain

→ PCG, §9 bosses, §11 SQs, §13 scene rows, seams).

The stack facts (Josh, 2026-07-15 — supersede anything older in the pipeline docs)

say "UE5" generically; the correction pass pins 5.8 and re-checks per-subsystem API surfaces).

Meshy and the other tools named in the existing build-pipeline docs. Any doc citing the

retired tools gets flagged and re-based during this review.

workstation: RTX 5090 (32GB), 9950X3D, 2×64GB (128GB), 4TB Gen5 (slot 1), 4TB Gen4 (slot 3),

8TB SATA (slot 4), slot 2 empty. Local 7B/3B available today; 14B-class local inference on

the 5090.

disjoint concerns (Fable = review/correction/contracts; Opus 4.8 1M Ultracode = volume

authoring under the corrected specs — the standing tiering policy).

The acceptance test (Josh, 2026-07-15 — the review is not done until these three chains

have complete, concrete, doc-carried answers)

1. The noise gate: how book processing filters noise and chunks-to-vector ONLY what is

relevant to region pages and canon — staging curation, per-source triage, relevance gating,

and retrieval-time filtering, stated end to end.

2. The world chain: how the pipeline downloads/creates the 3D earth maps, ALTERS them, then

generates and populates via 3D generation — and how asset/art/architecture creation hangs off

that — tool by tool, input contract by input contract.

3. The UE 5.8 chain: exactly what the engine build needs at EVERY step and how each step

picks its inputs off our schema and canon (the v3 spine graph, the region pages, the

registries) — no step left as a named-but-unspecified hop. Initial research landed 2026-07-15

at docs/pipeline_review/tech_research/UE_BUILD_AUTOMATION.md (the full build-automation

surface + a step-by-step chain draft + a residual-unknowns register) — raw material toward

this acceptance test, not yet critic-reviewed or ratified as the final answer; see also the

STACK_FACTS_QUESTIONS.md Q1 addendum for the Epic-first-party-MCP-vs-NeoStack reconciliation

point it surfaced.

The corpus under review (the survey set)

(commits f409620, 8fe3a14) + docs/BOOK_ACQUISITION_MINIMIZED.md (the purchased set).

tier doc (LIVING source — corrections land in place).

docs/COMBAT_ENCOUNTER_SYSTEM.md (boss-as-system + milestone pipeline).

generative-entity-system-vision — deterministic-first, on-device 14B reference, Nemesis diff).

The mission plan (phases; each lands + commits before the next)

gaps (anything consuming the old v2.5 shapes, prose blocks, or the retired T2 stubs), (2)

retired-stack citations (Meshy-class tools, anything superseded by Neostack/NVIDIA/5090-local),

(3) internal contradictions and unfinished sections ("not flushed out"), (4) the hardware

assumptions vs the real timeline. Deliverable: docs/pipeline_review/GAP_MAP.md.

derive: exactly which written-stack pieces are being moved away from, what Neostack AI's

Unreal workflow expects as input, which NVIDIA generator (and its input contract), the chosen

5090 asset-gen model. Deliverable: a short structured question set (one sitting, everything

batched — never dribbled) + the ratified STACK_FACTS section appended here.

consumers, exactly what a region page must carry (the join surface: sites/terrain → GIS/PCG;

personas/factions → NPC gen; SQ/beat hooks → quest wiring; scene rows → cinematics; the RCM

substrate links; the book-processing outputs it absorbs). Deliverable: the region-page

TEMPLATE v2 + its own gate (the anti-orphan rule: a template is not done until a checker

enforces it and DOC_MAP carries it).

COMPLETE — all six contract slots authored (docs/translation/ — UE5_Build, AssetGen [Meshy],

Cinematic, Audio [AIVA+ElevenLabs+SFX in one audio-family doc], plus Terrain as a bonus

split), every doc derived from the researched stack with spikes triaged, the audio lane

RESEARCH-CLOSED (tech_research/AUDIO_STACK.md). DONE 2026-07-15: (1) the three-critic

fresh-context pass (zero blockers; the fix-forward landed 9328174 — Cinematic's worked cases

corpus-corrected, the shared anchor_type mis-model fixed, schema §3d corrected, 4 spine

alignments); (2) the CONTRACT CORRECTIONS PASS landed on T1_Build_Pipeline_Contracts (the

internal restructure log's v1.1 entry itemizes it; filename label stays v1.0 for

cross-reference stability): §0 repo-native Translation-Doc roster (all six slots authored),

§1.2 terrain supersession note, §2 Meshy-retired banner (schema KEEPs binding), §3.4 AIVA

API correction, §5 executed-spine banner + §5.5/§6.5 example re-keys + BOTH path_bucket

enums → the ratified vocabulary, §6.2 Two Courts correction, §7.3 S19 → ratified WS

vocabulary, §8 built-pipeline note, §9 wholly-superseded banner, §10 creature-roster input +

tool-ordering supersession + two ADDED open items (the PROPOSED adaptive-audio fields =

Josh's schema call; the SFX path convention). REMAINING in Phase D: the full §9 re-base

(rides the Opus lane per OPUS_HANDBACK_POST_V3); the DRAFT→ACTIVE flips for the five

Translation Docs (recommendation: flip after Josh's read of this landing — the critic pass

ORIGINAL SCOPE: Re-base each doc on the v3 spine + the ratified stack facts;

retire dead stack references; flush out the unfinished sections; wire the contracts tier doc

in place. Fresh-context adversarial critic before any ACTIVE flip (the standing gate).

COMPLETION: finalize the processing structure (inbox → extraction → per-region substrate →

region-page absorption), sized for local-first bulk work (7B/3B now, 5090 later) with paid

tokens only on quality-bearing synthesis.

Operating notes

work the corrected specs unlock, PLUS the post-v3 canon work orders in

docs/OPUS_HANDBACK_POST_V3.md (runs in the parallel lane, disjoint files, own commits).

standing landing loop; the corpus spine_graph gate must stay green.

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