REALM_ANALYSIS_ART_PIPELINE_2026-07-27.md

supernatural/REALM_ANALYSIS_ART_PIPELINE_2026-07-27.md

THE ART-CONCEPT SELECTION PIPELINE — the director-led choice-board program for the realms

Read-only design research, 2026-07-27. Axis: the art-concept selection pipeline for the mythological

realms (Ch 66-74) and the fairy realm (Ch 75-77 + Prologue + Ch 1). Produced against the repo as it

stands; every current-state claim carries a file citation. This doc invents no world canon — it

designs a PROCESS that consumes canon and produces choice boards for Josh. It proposes one new

harness tool, one new artifact class (the realm art bible), four critic rubrics, and a rights

discipline that closes a gate the repo already opened and left unclosed.

Cited short-form used throughout:

---

0. The directive restated as a machine

Josh ordered a director-led concept process that presents him with CHOICES per realm and per zone:

the most accurate depiction of that world, the recommended enhanced version for playability and fun,

and surrounding recommendations for story, side quests, easter eggs, exploration, adventure, action,

and nearby important zones. Everything stays automated; decisions come from deep-dive fan-outs,

alternatives analyses, and adversarial critics across all lenses. His input is the PICK.

Restated as a machine with a countable unit of work:

Catalog — one row per zone with terrain, gameplay functions, ambient_threat, vril_density,

climactic_affordances, protected, area_bounds, build_priority — and RPT L102-127 declares seven

extension columns of which three are directly this program's outputs and are currently EMPTY by

declaration: care_tier, asset_routing, and environment_grammar_refs. The pipeline does not

need a new spine of its own; it fills a table that is already waiting.

rules its whole visual grammar, and only after that board is picked do its zone boards run under

the ratified grammar. Ruling a realm zone-by-zone would produce a realm assembled from nine

independently-good rooms that do not read as one world, which is the exact failure the realm-stratum

vision names ("realms must be REAL otherworlds, never slop").

factory reads. Boards are the deliberation record; the bible is the law.

The nine mythological realms in road order, read from the built spine rather than from any older

roster (docs/spine/CH_66..74.md L1): Ch 66 Atlantis, Ch 67 Pohjola / Sariola, Ch 68 the Apsu,

Ch 69 the Duat, Ch 70 Hades and Elysium, Ch 71 Xibalba, Ch 72 Naraka and Meru, Ch 73 Annwn, Ch 74

Yggdrasil. Then the fairy realm at Ch 75-77, and the Prologue plus Ch 1 built last of all. The

directive's build order and the spine's build order already agree, which means the program needs no

re-ordering argument — it inherits one.

---

1. The board — what Josh actually sees

1.1 The shape

One board per decision unit. Three panels, a marked recommendation on each, and a pick line. The

board is a self-contained page — offline-safe, no external fetches, theme-aware — delivered either

as a rendered file to the chat or as a published page, per §8 open item 1.

The board's own discipline mirrors the standing decision protocol rather than inventing a new one:

every panel carries options, a one-line rationale per option, a marked RECOMMENDATION with its

reasoning, and the STRONGEST OBJECTION to that recommendation. A board that presents images without

a recommendation is a bare question wearing pictures, and the protocol forbids it.

1.2 Panel A — THE ACCURATE DEPICTION

What the source tradition itself says the place looks like. Three options, each a plate sheet rather

than a single image: silhouette and architecture, palette and light, and one signature detail.

Each option carries four things and no fifth:

to. "The Duat as the Book of Gates renders it: register-banded, procession-oriented, the hours as

architecture" is a rationale. "Egyptian underworld, mysterious" is not.

harness/recall.py plus the region-filtered substrate — and the positive-control rule from memory

applies: a citation line that cannot resolve is a defect, and an empty citation line means the

option is not an accurate depiction at all, it is a guess with good lighting.

than re-derived per board.

The bar this panel is held to is authenticity, not reverence. The care doctrine at its bolder

calibration binds here in both directions: the panel must not scrub what the tradition actually says

(the Lords of Xibalba are the Popol Vuh's own antagonists and the board renders them as such), and it

must not composite across unrelated traditions to make a prettier plate (BPC §2.8, carried into

TD_ASSETGEN §6: "sacred architecture depicts per-region/era form, no compositing across unrelated

traditions").

1.3 Panel B — THE ENHANCED VERSION

The recommended enrichment for playability and fun. Three options. The governing rule, stated so it

can be audited rather than merely intended:

amplifies scale, light, density, motion, verticality, or legibility, and it contradicts nothing on

Panel A. An enhanced option that changes what the place IS has left the program; it belongs in a

canon brief, not on an art board.

the audit surface. A reviewer who cannot read those two fields off the board cannot tell

enhancement from invention, and neither can Josh.

that buys a climb route, luminance banding that buys wayfinding without a marker, density that

buys an ambush lane. An amplification with no play claim is decoration, and decoration is the

first symptom of the slop the AAA re-architecture exists to prevent.

Panel B is also where the binding physics doctrine becomes visible. Every enhanced option states the

ALTERED LAW in one clause — which real physical behavior vril bends here, and how the bend reads on

screen. The Apsu's option might read "buoyancy inverted at a depth boundary — the freshwater sea

holds its own ceiling, and things fall UP into it past the halocline." The point is that the clause

is falsifiable: a physics critic can check whether the image shows a coherent altered law or shows

physics simply abandoned (§5.3).

1.4 Panel C — THE SURROUNDING RECOMMENDATIONS

Not images. A text card set, five slots, because Josh asked for exactly five kinds of thing and a

board that blends them into prose forces him to do the sorting.

held, what the place is doing for the chapter it sits in.

(docs/proposals/quest_seed_queue/) with a housing ledger and a check_seed_housing.py gate, so

Panel C's quest cards mint into an existing lane rather than a new one.

hard puzzles for. An easter egg that gates progression is a defect by standing ruling.

Pillar 3 (FINALE §1: "No chapter is combat-absent... combat-absence is creatively fatal"), so this

slot always has content, and a realm board that returns an empty action slot has found a design

defect rather than a quiet place.

realm from a corridor of rooms into a place. This slot writes into the Zone Catalog's own

neighbor-facing fields rather than into prose.

Each card carries a provisional id in the target registry's own id space so the pick is an APPLY,

not a re-authoring pass.

1.5 The pick line

The whole point of the board is that Josh's return is short. The pick line is one line per panel:

under the water").

A fourth lane exists and is named on the board so it is never improvised: MORE LIKE THIS. Josh names

a direction, the board regenerates, and the regeneration cites the previous board so the reasoning

chain survives. Nothing about the pick line asks him to write a brief.

1.6 What may never appear on a board

players never see the verb taxonomy is about player surfaces, and a board is not one — but the

board's captions are the seed text for downstream surfaces, and taxonomy that enters here leaks

outward. Boards use natural language for the same reason the game does.

gate enforces this (§5.1) because a board with two recommendations is a board that failed to decide.

---

2. Generation-stack integration — what runs when

2.1 Concept art is a new slot, and saying so is the honest move

TD_ASSETGEN is the Slot 3 translation doc and it covers 3D GEOMETRY — the seven-category taxonomy,

mesh_tags, the five-state machine, the finishing pipeline. Its own scope line excludes voice,

terrain, editor, runtime, and OCR, each of which has its own doc. There is no concept-art slot. A

grep across docs/ for a concept-board or art-bible contract returns nothing but incidental

mentions, and DESIGN_GAP_REGISTER carries no art-direction entry. This program is therefore a NEW

translation-layer slot, and it should be declared as one rather than smuggled into the AssetGen doc

as an appendix — the AssetGen doc's authority is over meshes, and a concept-image lane inheriting its

constraints by implication is exactly the kind of unstated inheritance that rots.

The relationship between the two is one-directional and clean: the art bible this program produces is

an INPUT to TD_ASSETGEN §4.1's mesh_tags composition, specifically the style sub-field, which

TD_ASSETGEN §6 already names as the downstream audit anchor ("mesh_tags.style records the selected

style for downstream audit"). Concept decides; geometry executes.

2.2 The three lanes

LaneWhat it isWhere it may appearGating
R — REFERENCERights-cleared real imagery assembled into plate sheets. No model runs.Panel A (primary), Panel B underlaysUngated — runs today
P — PAINTOVERGenerated imagery composed over a Lane-R underlay, inheriting the reference's structurePanel B (default)GATED-5090 for volume
G — GENERATEDRaw local text-to-imagePanel B for fictional realms; Panel A only for wholly fictional realmsGATED-5090

Lane R is the one the repo is best positioned for and the one nobody has scheduled. docs/BOOKS_OWNED.md

L138 records that Josh owns the complete National Geographic run since 1888 and names its primary

value as VISUAL REFERENCE. That is a rights-encumbered but genuinely owned reference corpus covering

most of the arc's geography, and it is sitting unused by any visual lane. Lane R also draws museum

open-access collections, which are the single cleanest rights class available for the living-tradition

realms.

Lane P is the recommended default for Panel B rather than Lane G, and the reason is structural rather

than stylistic: a paintover inherits the reference's proportions, so the enhanced option is a

superset of the accurate one BY CONSTRUCTION instead of by a reviewer's judgement. That converts

§1.3's "reality is the floor" rule from a promise into a property of the process.

2.3 Model roster constraints, inherited not re-argued

TD_ASSETGEN §1's four ruled constraints bind this lane too, and three of them transfer verbatim:

carry a MAU cap, territory exclusion, revenue ceiling, or non-commercial clause is disqualified from

producing anything that reaches an art bible. Concept art that informs a shipped asset is arguably

more license-entangled than a mesh, not less, because style transfer leaves fingerprints.

load-bearing. This is what the previz_only rights value in §3.3 encodes at row level.

extends it across folklore creatures, dress, adornment, and sacred architecture. THIS PROGRAM

EXTENDS IT ONE STEP FURTHER, and the extension is the doctrinal contribution of this axis: the rule

covers realm ICONOGRAPHY of living traditions, not just built structures. The Duat's registers and

gates, Naraka's cosmological strata, Xibalba's house-trials, and Meru's axis are ICONOGRAPHIC

SYSTEMS with their own internal grammar, and a model's hallucinated guess at that grammar carries

the same class of error as a hallucinated temple — it gets a living tradition's own picture of its

own cosmos wrong, at AAA fidelity, in a shipped product.

2.4 The costume rule, extended to the board

TD_ASSETGEN §6's photogrammetry-costume rule says a generated image run through a reconstruction

pipeline produces an asset "wearing photogrammetry's credibility as a costume." The board has a

direct analog and it needs stating because a board is a persuasion surface: a generated image placed

on PANEL A wears accuracy's credibility as a costume. Josh reads Panel A as "this is what the place

actually looks like." An image that is a model's guess, sitting under that header, has told him

something false in the one place he is relying on the pipeline to be true.

So the rule is flat and takes no judgement call: Panel A is Lane R for every realm whose source

tradition is living, and Lane G is permitted on Panel A only where the realm is wholly a literary or

game-canon invention. Atlantis qualifies (the spine's own Ch 66 header calls it Plato's literary

invention and carries an anti-fringe firewall). The fairy realm qualifies (FINALE §6: "the fairy

realm the game's own L3 fiction"). Nothing else on the road does.

2.5 Which realms carry which load

The nine realms do not share a care profile, and pretending they do produces either nine

over-constrained boards or nine under-constrained ones. The split that matters is whether a LIVING

community holds the tradition, because §17.1 REPRESENT_WITH_CARE is collective and living-facing.

heritage plus a very large real monumental corpus), Naraka and Meru (Ch 72, living Hindu, Buddhist,

and Jain cosmology — the sharpest case on the road, because these are not historical cosmologies but

currently-held ones), Xibalba (Ch 71, living Maya communities and a source text, the Popol Vuh, that

is itself a living cultural possession).

reference-first on quality grounds: Pohjola / Sariola (Ch 67, the Kalevala — note the Finnish

national-epic status makes this closer to the living column than a casual read suggests), the Apsu

(Ch 68, Mesopotamian), Hades and Elysium (Ch 70), Annwn (Ch 73, Welsh — and Welsh tradition has a

living claim, so this sits closer to the elevated column than to Hades), Yggdrasil (Ch 74, Norse —

with the standing caution that Norse iconography carries a modern extremist-appropriation tail that

the vril firewall memory is already alert to).

The routing is not a per-realm judgement made fresh each time. It is read off the Zone Catalog

care_tier column, which RPT L120 already defines as derived from the protected flag plus any ruled

care-register overlay. Populating that column for the realm zones is a prerequisite of the first

board, and it is ungated desk work.

---

3. The rights discipline

3.1 The gate the repo opened and left open

TD_ASSETGEN §6 opens the capture-rights gate explicitly and closes nothing: "Before any

reference-photogrammetry input reaches modeling, it needs a per-site capture-rights check —

provenance and permission, not just realness — parallel to the use_class discipline already

governing the book-processing substrate lane." DROPIN §7 item 4 raises the same gate as an open item

for director ruling. docs/BUILD_PLAN_END_TO_END.md HOP 10 states it as MISSING outright: no

reference_image_ref, capture_rights, reference_source, or use_class field exists anywhere in

the asset lane. docs/PIPELINE_READINESS_ROADMAP.md L66 records the same absence and its consequence

— the agent cannot distinguish a rights-cleared photo from a prohibited one, or from AI-generated

input.

Three docs name the same hole. The concept lane is where it becomes urgent, because concept art is

where reference imagery ENTERS the pipeline. Closing it here closes it upstream of every downstream

consumer.

3.2 The precedent to copy rather than invent

The book-substrate lane's use_class field is the working precedent both TD_ASSETGEN §6 and the

build plan point at. The pattern is: a per-row enum recording what may legally be done with this

source, checked by a gate, refusing to let an unpermitted class flow into an output. Copy the shape.

3.3 The rights vocabulary

Every image on every board carries exactly one value:

the living-tradition realms, and the reason Lane R is viable at all for the Duat.

image does not by itself clear a site-access or cultural-protocol question, so this class still

routes through the cultural-authenticity critic.

Panel A only for invented realms.

Usable internally for thinking. NEVER promotable to an art bible, and never on Panel A.

Two rules make the vocabulary do work rather than decorate:

the promotable set. previz_only is the one value that can never promote, which is precisely the

behavior TD_ASSETGEN §1's constraint 2 demands and which nothing currently enforces.

commercial capture or commercial 3D reconstruction regardless of who took the photo. The per-site

check TD_ASSETGEN §6 asks for is a second field, not a substitute for the first, and the realm

program is the natural place to populate it because the realm boards are the first work that

systematically assembles real-site reference at scale.

---

4. The selection loop and the ratified art bible

4.1 The loop

describes — architecture, palette, light, materials, scale, inhabitants, the place's own internal

logic. Grounded in the region-filtered substrate plus the spine node plus the RCM section. Output is

a research pack, cited, that both Panel A and Panel B draw from. This is the largest single body of

work in the program and it is entirely ungated.

distinct readings rather than three renderings of one idea. The standing decision protocol's

requirement that the option set include the already-applied case has a direct analog: where a realm

already has an authored visual description in its spine node, that description is an option on the

board, named as the incumbent.

board; Josh never sees a board that failed a floor.

that moment. The pick line is quoted verbatim in the bible so the ruling is never paraphrased —

this matters because of the standing lesson that a subagent writing a user ruling into canon

reads as fabricated authority; the ratification write is a DIRECTOR-level apply, not a subagent one.

4.2 The art bible

One per realm, docs/art_bibles/REALM_<node>.md. It is the ONLY artifact the asset factory reads,

which is what keeps the boards from becoming a second source of truth. Contents:

without placement rules produces nine realms that all use their palettes the same way.

realm's light does that Earth light cannot, stated as an altered law rather than as an effect.

behavior so the physics critic has something to check.

whether a realm reads as itself from a hundred meters away.

Catalog's meters_per_vertex rung (RPT L114) so the art call and the geometry budget agree.

mechanical read. This is the field that carries the directive's "physics blending perfectly and

harmoniously with metaphysical lore," and it is the field that hands to T0_Environment_Grammar_Registry.

otherworld, since motion is where "this place obeys different rules" is felt rather than seen.

as specific prohibitions rather than as a quality aspiration.

4.3 Downstream consumers, so nothing orphans

The anti-orphan rule is standing: every artifact needs a consumer AND a DOC_MAP row. The bible's

consumers, each already existing:

sharpest fit in the repo for the physics doctrine: its columns include physical_properties_array

(EGR header: mass:none;thermal_state:hot;conductivity:low), elemental_affinity,

tactical_affordance_class, attack_verbs, enemy_usable, and protected_care_ref. A realm's

altered law lands there as MECHANICS, not as an art note — which is what stops the realms from

being beautiful backdrops with ordinary combat in front of them.

environment_grammar_refs (populated by the scatter), and asset_routing (blocked on an enum

ratification, §8 item 4).

rule) and Section 17 (UE build specifications, environment areas, vegetation palette, performance

budget).

the frames against the ratified bible rather than against a reviewer's memory (§5.5).

---

5. The critic layer

5.1 The deterministic pre-critic

Before any model reads a board, a deterministic tool gates PRESENCE and SHAPE and never content. This

is not a new idea to invent — READCAP is exactly this tool for the capture lane, and its docstring

states the discipline precisely: "It gates PRESENCE and SHAPE only... and it NEVER judges image

CONTENT. Every must_show / must_not_show verdict belongs to the image critics downstream, never here."

Propose harness/read_boards.py, built on READCAP's interface contract rather than a fresh one: a

board manifest joined against a board-expectations file, emitting a critic worklist and a scorecard.

What it asserts:

transfer: a promised-but-missing plate is a structural FAIL; an expectation with no plate is a WARN.

is the mechanical form of the costume rule and it is the single most valuable tooth in the tool.

force here: the tool must be self-tested with a fixture that MUST fail, or it is not armed. READCAP

already carries a --self-test mode with must-fire and must-not-fire fixtures and exit code 3 for

broken teeth; copy that shape exactly.

5.2 The cultural-authenticity critic

Fresh-context, per board, calibrated in BOTH directions per the care doctrine's ruled method. It asks:

demon is a BOSS. Xibalba's Lords, Naraka's wardens, and the Apsu's sea are the arc's marquee

antagonist material, and a board that renders them as scenery has committed the exact defect the

rated-R memory exists to catch.

deity-name on a mortal NPC or a revered figure cast as villain?

option TIMID? Care-tier inflation is a defect by ruling. A critic that only ever pushes toward

restraint will produce nine reverent, empty realms, and the doctrine names that as the more costly

error, not the safer one. The critic reports timidity findings in the same list as insensitivity

findings, with the same weight.

5.3 The physics-coherence critic

The critic that makes the directive auditable. The binding doctrine is that magic works by ALTERING

real physics; both sides wield both. The test per image:

what direction. If the only available explanation is "it is magic," the image fails.

image whose alteration has no vril-consistent cause is inventing a second magic system by accident.

buoyancy inverts, things that are not the hero object also fall upward; if light propagates along

a different path, shadows agree. Consequence-consistency is what separates a real otherworld from a

filter applied to a photograph, and it is checkable from a single frame.

physical_properties_array and a tactical_affordance_class. A law that cannot become an

affordance is a law the player can only look at, and the realm-stratum vision's whole complaint was

realms that are looked at rather than played.

5.4 The AAAAA de-slop critic

Reads the board against reference-title bars and against the repo's own named slop-modes. The

visual translations of the recorded slop list:

road without anyone noticing, the board has failed. This is the single strongest test available and

it should be run literally, by presenting plates without labels and asking the critic to assign them.

consecutive realms of reverent gloom is the default failure of a mythological-realm road, and Annwn

in particular is an otherworld of PLENTY (the spine's own Ch 73 title says so) whose board rendering

as another dim underworld would be a canon error and an art error at once.

visual signatures — depopulated establishing shots, armies without faces — and the critic checks

for them on Panel C as much as on the images.

throne-room approach is the most common generative-image tell.

5.5 The reveal-discipline critic

The subtlety here is worth stating precisely, because getting it wrong in either direction costs

something. The reveal ceiling protects the PLAYER. Josh is the author, and boards are not player

surfaces, so a fairy-realm board being built while the realm road is still in progress leaks nothing.

What the critic actually guards is the DOWNSTREAM ROUTING of ratified art:

a chapter before 66. Fairy-realm iconography appearing in an early-chapter vision, a loading still,

or a marketing surface is the leak that matters.

(harness/check_grand_sage_silence.py) plus a reveal-discipline gate and a baseline

(docs/reveal_discipline_baseline.json). The art lane registers with those rather than inventing a

parallel check.

name-forbidden (HL_0099), no Grand-Sage hint appears before Ch 76, and the fairy-realm-script

identification is held until the Ch-8 reveal (docs/proposals/AMULET_PROVENANCE_BRIEF.md §2.6 via

PRES §0.B.1a). A board that renders the realm's script legibly in an early-chapter context breaks a

hard line with a picture, and no text gate would catch it.

5.6 Verdict shape

All four critics return the normalize-doc verdict shape already standard in this repo: GO /

GO-WITH-FIXES / NO-GO, ranked findings, per-finding evidence. NO-GO regenerates the board. Fixes are

applied and the board re-runs the deterministic pre-critic before delivery. The critics run at the

opus tier per the standing model-seat ruling; the deterministic tool is zero-token and runs anytime.

---

6. The pre-5090 and gated split

6.1 What runs today, with no GPU

This is the honest and encouraging part of the answer: most of the program is ungated.

actually describes. This is the long pole and it needs no hardware.

self-test fixtures, the art bible template, and the four critic rubrics.

been carrying as open.

reference corpus. This produces REAL Panel A content today.

uses.

6.2 The intensity ladder, which is ungated and which the directive specifically asks for

Josh asked for realms that "strongly build the intensity toward the end-arc fairy realm." That is a

ladder decision across nine realms, not a property of any one board, and it can be ruled before a

single image exists. The recommendation is that every board carries an INTENSITY SLOT showing where

its realm sits on the climb and what the two adjacent realms do, so Josh is ruling the ladder as he

rules the rooms. The topography discipline the repo already runs (peaks, valleys, no three

consecutive, mode variety — FINALE §5) is the mechanical analog and the ladder should be checked

against it, since a visual crescendo that fights the pacing topography will lose to it in play.

6.3 What is gated on the 5090

and any per-zone board beyond the realm establishing board.

category, and TD_ASSETGEN §5 already specifies that protocol including a real-world-architecture

test whose stated purpose is to CONFIRM the never-raw-generate call rather than to greenlight it.

6.4 The honest pre-hardware board

A pre-5090 board is a real board Josh can rule on, and it should be shipped rather than deferred —

but its Panel B is a REFERENCE MOODBOARD rather than a generated concept set: existing rights-cleared

art and photography that demonstrates the amplification DIRECTION, with the same amplified and

stayed-true labels. This is weaker than a generated concept and the board says so on its face. What it

buys is that the palette, light doctrine, silhouette grammar, altered law, and the whole of Panel C

can all be ruled before hardware arrives, so the 5090's image time is spent rendering ratified

directions rather than exploring open ones. That is the correct order of operations regardless of

hardware, and the constraint is doing the program a favor.

---

7. Build order and the pilots

Road order, per the directive and the spine, which agree: the nine mythological realms Ch 66 through

Ch 74, then the fairy realm, then the Prologue and Ch 1 last of all.

Two pilots are recommended rather than one, because the road has two genuinely different problems and

piloting only the easy one would validate half the machine.

the spine node's own header), so it exercises the complete board, both generation lanes, the pick

loop, the art bible, and the scatter, with no capture-rights blocking. It also carries a real

anti-fringe duty (the hyperdiffusionist-racism firewall the node names), which exercises the

cultural-authenticity critic in its most interesting mode — a fictional realm with a live care

obligation attached to its RECEPTION history rather than to its content.

and iconographic corpus, elevated care, and a Panel A that must be assembled from rights-cleared

reference rather than generated. If the machine works here it works everywhere on the road, and the

rights vocabulary gets tested against a realm where the answer actually matters. Ch 69 is also the

materialized-argument node, which means its board must serve the boss class whose shared visual

grammar is already RULED (TD_ASSETGEN §2, D-MATARG-LOOK: inscription-made-flesh, per-boss regional

condemned-register skin, shared hollow-core tell, and the bound that a construct never wears a

culture's sacred beings). That ruled grammar is a live constraint the board must respect and a

worked example of exactly what an art bible looks like when it is done right.

7.1 Engine versus canon

Per the standing engine-reuse directive, declared up front so game #2 harvests it:

self-test fixtures, the rights vocabulary and its promotion gate, the four critic rubric SHAPES, the

art bible TEMPLATE, and the pick-to-ratify-to-scatter loop.

vril-physics signatures, the care-tier assignments, and the reveal-window constraints.

---

8. Open items — director ruling or Josh ruling, with recommendations

Every item below is presented in the standing protocol's shape rather than as a bare question.

1. BOARD DELIVERY MEDIUM. Options: a rendered file sent to chat; a published self-contained page; a

local HTML file opened directly. RECOMMENDATION: a self-contained page, because boards are

image-dense, need side-by-side comparison, and benefit from persisting at a stable link across the

nine-realm program. STRONGEST OBJECTION: publishing puts unratified concept art at a URL, and

fairy-realm concepts in particular are the most spoiler-dense artifacts the project will produce —

which argues for local files for the fairy realm specifically, and pages for the rest.

2. OPTIONS PER PANEL. RECOMMENDATION: three. It is the count at which a genuine choice exists and at

which each option can still carry a full rationale, citation, and objection. STRONGEST OBJECTION:

three under-samples a nine-realm program's visual space, and a fourth "wildcard" option costs

little to generate — the counter-recommendation of three plus an optional wildcard is defensible

and Josh may prefer it.

3. WHETHER PANEL A MAY BE GENERATED FOR INVENTED REALMS. RECOMMENDATION: yes, and this is the pilot's

whole point — Atlantis and the fairy realm have no tradition to be inaccurate about. STRONGEST

OBJECTION: it creates a two-rule board where a reviewer must know which realm they are looking at

to know which rule applies, and rules with exceptions decay. The mitigation is that the rule is

keyed on the care_tier column rather than on reviewer memory, so the deterministic gate applies

it, not a person.

4. THE asset_routing ENUM RATIFICATION. RPT L124 declares the column and states it stays empty

"until the §2.6 / AG7.1 asset_category/asset_routing enum is ratified (an unratified enum value

would be invented canon)." This program needs it and is the natural forcing function. It is a

small, gate-safe ratification and it currently blocks a declared column from ever being populated.

5. THE CONCEPT-ART MODEL ROSTER. A license read is owed before any Lane G or Lane P generation,

identical in kind to TD_ASSETGEN §5's pre-flight. RECOMMENDATION: fold it into the existing 5090

benchmark-day pre-flight rather than scheduling a separate exercise, since that pre-flight is

already specified as a cheap GPU-free license re-check.

6. WHETHER THE CONCEPT LANE BECOMES ITS OWN TRANSLATION DOC. RECOMMENDATION: yes —

docs/translation/T99_Translation_ConceptArt.md, sibling to AssetGen, Terrain, Cinematic, Audio,

Combat, and UE5_Build, subordinate to the CVD on canon and to Contracts on schema, owning the

concept layer none of them own. STRONGEST OBJECTION: a seventh translation doc is a maintenance

surface, and the alternative — a section inside TD_ASSETGEN — is cheaper. The counter is that

TD_ASSETGEN's authority is over meshes, and a concept lane inheriting mesh constraints by

implication is precisely the unstated inheritance that goes stale.

---

9. The three findings worth carrying forward on their own

photogrammetry pipeline from generated input. The identical error is available one step earlier and

is easier to make: a generated plate on Panel A tells Josh something false in the one place the

board promises truth. Stating it, and enforcing it mechanically on the care_tier column, is the

highest-value single rule in this design.

covers real cultural ARCHITECTURE. The realms are made of cosmological grammars — the Duat's

registers and gates, Naraka's strata, Xibalba's houses, Meru's axis — which are living traditions'

own pictures of their own cosmos, and which a model will confidently hallucinate. This is a

genuine extension of a ruled constraint into new territory, and the realm program is the first work

that hits it.

T0_Environment_Grammar_Registry already carries physical_properties_array, elemental_affinity,

tactical_affordance_class, attack_verbs, enemy_usable, and protected_care_ref. Routing each

realm's altered law through an EG_ row is what stops nine gorgeous realms from having ordinary

combat in front of them, and it is the concrete mechanism behind the directive's "physics blending

perfectly and harmoniously with metaphysical lore." An art bible field that cannot become an EG_ row

is a field describing scenery.

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