CONCEPT_ART_PRODUCTION_PROGRAM.md

assets/CONCEPT_ART_PRODUCTION_PROGRAM.md

THE CONCEPT ART PRODUCTION PROGRAM

CANON SUBORDINATION — this document is PROPOSAL-TIER: it serves canon and never outranks it.

The canon: the CVD · the T1 foundation docs · the T0 registries (registries/) · the spine

(docs/spine/CH_*.md) · the region pages (_source/02_Tier_2_Region_Pages/). Authority order:

docs/DOC_MAP.md § 0.

Canon served: docs/spine/CH_*.md § ## Asset anchors + § ## Mechanical anchors (79 nodes) ·

registries/T0_Chapter_Index [ACTIVE v1.2] · registries/T0_Region_Index [ACTIVE v1.2] ·

registries/T0_Creature_Roster · registries/T0_Character_Index [ACTIVE v1.1] ·

registries/T0_Boss_Encounter_Registry · registries/T0_Weapon_Registry ·

registries/T0_Ability_Tree_Registry [ACTIVE v0.1] · registries/T0_Familiar_Registry ·

_source/02_Tier_2_Region_Pages/*.md § Zone Catalog.

READ THAT CANON FIRST — open it and derive from it before you build anything from this document.

If this document disagrees with canon, CANON WINS and this document is the defect — fix the

document, never the canon. Nothing here is applied until it is ratified into canon.

Status: PROPOSAL-TIER. v1.0, 2026-08-02.

It answers Josh verbatim: *"build the concept art (hundreds for the region and thousands for all

the assets architecture fauna npcs etc in the region, and thousands more for everything the player

will walk through and experience, and thousands more for abilities, and thousands more for weapons,

and thousands more for characters and companions and npcs, and thousands more for equipment), not

all is needed at once and you develop those pages and I approve them or make comments on how to

adjust or modify."*

0. WHAT THIS IS, AND WHAT IT DOES NOT RE-RULE

docs/proposals/CONCEPT_ART_PROGRAM.md (1,840 lines) answers WHAT art exists and how much — the

43-class density model, the three-panel board shape, the licence routing. It is the TAXONOMY and it

stands; nothing here re-derives a single one of its counts.

This document is the other half: **HOW ~17,900 plates are produced, stored, reviewed, and kept

honest, by one person with one GPU.** Six questions, six sections, every number traced.

It does not re-rule the board shape (CONCEPT_ART_PROGRAM.md §2.1), the board medium

(docs/art_boards/README.md "The board medium is RULED"), the six-value rights vocabulary

(docs/art_boards/README.md "Conventions the gate enforces"), the licence tiering

(docs/pipeline_review/tech_research/PIPE_ART_2026-07-29.md §1), or the five-state generation

lifecycle (docs/generation_lifecycle.json enum). It **uses all five and invents no parallel of

any of them.** Where it proposes something new the text says PROPOSED.

The failure this program is shaped around. On 2026-08-01 a session built BRD_0001 from the two

program docs and never opened docs/spine/CH_02.md, whose ## Asset anchors block at line 226 had

already typed and id'd the board's entire contents. It passed the structural gate and four

adversarial critics. harness/check_canon_derivation.py (GATE 39) closed the derivation hole after

the fact. **Every mechanism below is designed so that failure is not merely caught but

uneconomic** — the cheapest path through this program is the one that starts at the canon node.

---

1. THE COUNTED VOLUME

Counted from the live registries on 2026-08-02, not from any document's summary.

**Multiplied to plates by docs/concept_art/density_model.csv (43 classes, emitted by

harness/concept_art/emit_density_model.py, never typed):**

Two of those counts are ceilings, not rows, and the program must not pretend otherwise. The

density model sizes creatures at 420 and named entities at 1,000+ against live registries holding

150 and 164 — that is ~2,900 plates budgeted against rows that do not exist yet

(CONCEPT_ART_PROGRAM.md §6.2 item 1 states this openly). The production consequence is stated in

§5: you cannot generate art for a row that has not been authored, so registry population is a

hard predecessor, not a parallel track.

1.1 The number that actually sizes the work: REVIEW UNITS

Plates are not the review unit — a SHEET_4 boss is four plates and one decision. Dividing each

class's plate count by its art_unit arity gives the number of things a human must actually look at:

cheap_or_expensive is a live column of density_model.csv; the split is the model's own, not a

judgement made here. 20.9 % of all plates are parametric — they are produced by applying a rule

that one board sets (the rarity ladder, the per-element VFX law, the weapon grade ladder). Those do

not get looked at one by one, and §6 says exactly what they get instead.

---

2. THE UNIT OF REVIEW — THE LEAF, capped at 12 rows

Not one giant board. **A LEAF: one canon anchor line-set, at most 12 entity rows, one page, one

scroll.** A leaf is the atomic reviewable, and its cap of 12 is measured, not chosen.

2.1 Why 12 — the measured half

Across all 80 spine files, every bullet under ## Asset anchors and ## Mechanical anchors that

carries a canon id was counted:

No anchor line anywhere in the spine names more than 12 ids. So a cap of 12 can always split a

node's anchor section on a LINE boundary and never cut a line in half — the leaf boundary and the

canon boundary are the same boundary, at every node, by measurement. A cap of 10 would split

CH_02.md's creatures_and_fauna line (7 ids) away from nothing, but a cap of 8 would eventually

cut one; a cap of 16 buys nothing the canon grain uses. **12 is the smallest cap that never cuts a

canon line.**

And because the median anchor SECTION carries 9 ids, **most nodes' entire ## Asset anchors block is

one leaf.** The reviewable unit and the canon unit coincide without either being bent.

2.2 Why 12 — the review-time half (MODELED, not measured — the parameters are named)

Assumptions, stated so they can be argued with or replaced by a measurement after the first ten

leaves: an approve is ~12 s (look, recognise, tap), a comment is ~45 s (look, decide what is

wrong, type it), and ~20 % of rows draw a comment. Mean 18.6 s/row.

Josh's phrase is *"scannable on a phone in a few minutes."* 12 lands at 3.7 minutes. The row count is

the lever, not the plate count: a leaf of 12 boss rows is 48 plates and still 12 decisions.

The total human cost, which is the number that decides whether this program is real:

142 leaves for the entire Prologue→Ch 13 slice ≈ 8.8 hours of Josh, and ~833 leaves ≈ 52 hours

for the full game — plus ~20 law leaves (§2.3) that cover the other 2,851 parametric units.

Spread across the build that is roughly one leaf a day. The GPU is not the bottleneck; this is —

which is why §5 orders the work by what unblocks the build rather than by what is cheap to render.

2.3 The three leaf kinds

The ANCHOR LEAF is the load-bearing one and it is the direct structural answer to the 2026-08-01

failure: **a leaf cannot be authored without opening the anchor section it is bound to, because the

anchor section is what enumerates its rows.** There is no leaf-shaped thing you can build out of

CONCEPT_ART_PROGRAM.md.

2.4 What a leaf page shows, per row

One row = one entity = one decision. Recommended plate at full page-width; the two alternates as a

collapsed strip; the entity's canon id and its [SRC:] anchor visible as text; one approve control

and one comment field. Below the rows, unchanged from GATE 39 D5: **the residue block — every id

this leaf declares pending, with its reason and its owning registry — rendered on the page's own

face**, so a leaf that realized nothing is visibly a leaf that realized nothing.

---

3. THE ID AND STORAGE SCHEME

3.1 Ids — extend, never mint a second key

plate_id (PLT_0001_A1_1) and board_id (BRD_0001) are landed keys in

docs/art_boards/PLATES.csv / MANIFEST.csv. They stay. Nothing here mints a parallel id space.

What is added is a path, because the landed asset_path column has no ruled form and 17,876

free-form paths is not a scheme.

3.2 The path form and the canon-id law

AMENDED IN PLACE 2026-08-03 (thirty-first sitting — the pipeline-hardwiring mandate, applied

through the four-lane compatibility pass). The path form below is unchanged and is exactly what

harness/check_art_pointers.py enforces. What this amendment adds is the LAW that fills the second

segment, the token grammar that carries the derivation onto the row instead of losing it at the CSV

boundary, the column set the row needs, the stage discipline, and the render-tier axis. Every clause

here was forced by a measured defect: the 2026-08-02 lane generated 76 plates whose second segment

held a prompt-pack kebab slug, so GATE 39 moved by nothing and no dispatcher could find a plate for

any canon row. Nothing below re-rules the path FORM — the form was right; the VALUES were not.

$HUMANITY_ART_ROOT / <class_id> / <canon_id> / <plate_role>__<hash12>.<ext>
                     D1-15        CR_0001      silhouette__9f2a1c4b7e03.png

• <class_id> — the density_model.csv class. Files sort into the same 43 buckets the schedule

budgets in, so "how much of D1-15 exists" is ls | wc -l and never a query.

• <canon_id> — the canon row the plate realizes. The path itself carries the derivation. A

plate that cannot name a canon id has nowhere to be written.

• <hash12> — the first 12 of generation_prompt_hash, which

harness/generation_prompt_hash.py already computes over a CLOSED payload **including the model's

licence class**.

That last choice is the whole design. Because the hash covers the prompt, the seed, the sampler

config, the tier and the licence class, *any* regeneration — a comment-driven re-prompt, a seed

change, a model demoted from SHIPPABLE to PREVIZ_ONLY — produces a different filename. So:

• a regeneration physically cannot overwrite the plate it replaces;
• docs/art_boards/README.md's "Rejected outputs … are kept, never deleted" stops being a habit and

becomes a property of the naming scheme;

• and the licence self-enforcement harness/generation_prompt_hash.py was built for now reaches the

filesystem, not just the row.

3.2a THE CANON-ID LAW — the class selects the registry, the canon id is a live key in it

AMENDED IN PLACE 2026-08-03 (director ruling R-G, the art final build). The sentence below used

to read "Nothing else may occupy segment 2 — not a prompt slug, not a beat id, not a site name, not a

title." That wording contradicted this section's own armed tooth: harness/check_art_pointers.py

P9 has always resolved zone: (against build/zones/<page_slug>.csv), beat: and scene: (against

docs/spine/CH_NN.md) as PRIMARY tokens, and site: through the class's key registry — so a law that

banned a beat id from segment 2 was banning something the gate lawfully admits, and the first row that

needed one (the landed fire-night row, D1-03/CH02_B02) had to be declared a deviation instead of a

key. The FORM is unchanged. What is corrected is which token families may be the PRIMARY.

SUPERSEDED: the old prohibition on beat ids in segment 2. STANDING: the prohibition on a

prompt slug and on a title, which is the defect the section was written for.

<class_id> SELECTS a registry. <canon_id> is a LIVE PRIMARY KEY — a key that a named repo source

resolves, which is exactly the set check_art_pointers.py resolve_token() implements:

• the class's own key_registry / key_column (entity:, boss:, insc:, site:, law:) — the

default and the majority case;

• zone:<zone_id>@<page_slug> — resolved against build/zones/<page_slug>.csv, page-qualified

because a zone_id is unique only within its page. Segment 2 is the bare zone_id;

• beat:CHNN_BNN and scene:SC_CHNN_<slug> — resolved against docs/spine/CH_NN.md's own

§10 beat lines and §13 scene-anchor table. A staging or key-frame plate whose subject IS a beat

keys on the beat; a beat with no §13 scene anchor may never have one minted for it to key on,

because the spine is what declares scene anchors and an invented SC_ id fails P9 by construction.

A prompt slug and a title remain banned, and so does any token no source resolves. The refusal at

§3.2b is untouched: a row with no live primary key is NOT GENERATED.

• The binding is DATA, not prose. docs/concept_art/density_model.csv carries key_registry and

key_column on all 43 class rows, so the class-to-registry rule is derivable rather than

remembered. key_registry is the docs/generation_lifecycle.json enrolment key

(T0_Creature_Roster/Sheet1) and key_column is that registry's own primary-key column

(creature_id) — the same pair generation_lifecycle.json already declares under id_columns

for the twelve enrolled registries. It is a re-point of live data, not a new id space.

• Two classes were REPOINTED because their old ruled_source named a cell with no primary key.

D1-16 folklore beings moved off T0_Chapter_Index.folklore_entities_in_chapter — a

semicolon-joined token list on a NODE registry — onto T0_Creature_Roster/Sheet1 creature_id,

because CR_0151 (the Polo) and CR_0010 (Ebu Gogo) are the live rows those tokens name. D1-03 and

D1-04 layouts moved off "spine 3b site rows" onto the ZONE key: build/zones/<page_slug>.csv

zone_id, which is the spine §3b site_id carried verbatim by harness/regen_zone_catalog.py.

• Three declared sentinels, and no fourth. (law) with key_column art_bible_id is the LAW LEAF

(§2.3) — a class whose subject is a rule, not a place, keys its plates on the art-bible artifact

id and never on a fabricated canon row; the two law leaves are the LANGUAGE_BOARD classes, D1-28

Hollow surfaces and D1-34 UI iconography and rarity language, and art_unit == LANGUAGE_BOARD is

the machine discriminator so the set cannot drift by opinion. build/zones/<page_slug>.csv is the

ZONE key, page-qualified because a zone_id is unique only within its region page. (none) means

no live primary key exists ANYWHERE, so the class cannot dispatch at all — D1-12 non-boss enemy

classes is the only one, and its home is DESIGN_GAP_REGISTER gap 1.

• A class whose subject genuinely spans two live registries declares BOTH, pipe-joined in the same

order in both columns (D1-20 minerals and gather nodes; D1-37 scripts and inscription spine). The

plate row's own canon_registry then says which of the two that plate keyed on, so the ambiguity

is resolved per row rather than left in the class.

• Segment 2 is the PATH-SAFE form of the primary token: BE_0002.P2 is written BE_0002_P2 in the

path while the untouched token stays on the row in canon_ref. A filesystem must never be the

reason a canon key changes shape, and a canon key must never be the reason a path breaks.

• All 43 pairs were resolved live against registries/** and build/zones/*.csv when this

amendment landed — 0 unresolved, with a positive control run on a deliberately wrong registry and

a deliberately wrong column so the zero is a real zero and not a dead search.

3.2b THE MULTI-ID RULING — three dispositions, and the refusal

"One row = one entity = one decision" (§2.4) is not advice; it is what makes the path form

expressible. A row that names more than one canon id takes exactly one of three dispositions.

• SPLIT. N peer entities at the same grain become N rows. A "flores ambient fauna study sheet"

naming CR_0003, CR_0004, CR_0005 and CR_0006 is four rows at D1-15/CR_0003, /CR_0004,

/CR_0005, /CR_0006 — which is what D1-15's own formula (species x 3) already asks for.

• PRIMARY + REFS. One entity seen through a relation is filed under the class whose registry owns

THE VIEW, and every other id rides canon_id_refs. An antagonist operative who is also a named

character and a boss is D1-11/NW_0001 with canon_id_refs CHAR_0008|BE_0002 — and he

separately OWES a D1-09/CHAR_0008 portrait and a D1-13/BE_0002 boss sheet. Three classes, one

man, three plate sets; never one plate wearing three hats.

• LAW LEAF. A plate whose subject is a rule applied hundreds of times (a material palette, a

style register, the Hollow surface language, the rarity language) keys on the art-bible artifact

id under (law). It names no canon row because it realizes none.

• THE REFUSAL, which is not a fourth disposition. A row with no live primary key is NOT

GENERATED. It goes to its board's coverage block as pending with its reason and its owning

registry, and the composer refuses the work order. This is the tooth that would have stopped the

four unmintable rows of 2026-08-02 (a realm object routed to a registry holding one row, a carved

mark with no home anywhere, a beat id, and a hard-line id that is a rule and not a thing).

3.2c THE canon_ref TOKEN GRAMMAR — the derivation, carried on the row

canon_ref is a pipe-joined list of prefix-typed tokens. The FIRST token is the PRIMARY, and the

path's second segment is its path-safe form. The polymorphic-reference precedent the repo already

carries is T0_Dialogue_Line.speaker_ref and T0_RNG_Drop_Table.linked_entity_id; nothing new is

invented here except the prefix set.

• entity:CR_0001 · entity:CHAR_0008 · entity:NW_0001 — a registry row in an id family.
• zone:mount_inerie_cone_country@flores_island — a zone row, page-qualified.
• beat:CH02_B06 — a spine §3c beat. Mandatory on any scene or quest-moment plate, because the beat

row is the only key that resolves both the level and the sequence slot.

• scene:SC_CH02_polos_offer — a §13 scene anchor, when the beat carries one.
• boss:BE_0002.P2 — a boss encounter, PHASE-qualified wherever the phase breakdown names more than

one phase, because two phases can be two different zones with two different tells.

• site:VS_CH02_001 · insc:INSC_77H_CH02_FRAGMENT — vril site and inscription spine.
• law:material_palette_and_style_register@flores_island — a law leaf.
• pending:<reason-slug> — the only legal value for a row with no canon target, and it is illegal

on a row whose generation_status is complete. A pending token obliges a matching pending row

in the owning board's coverage block, in the same commit.

AMENDED IN PLACE 2026-08-03 (director ruling R-B). pending is deliberately NOT a member of

harness/read_boards.py CANON_REF_PREFIXES, and the two do not contradict once the SCOPE is

stated, which it was not: CANON_REF_PREFIXES is the LANE-G/P prefix set. T-CITE-REQUIRED

inspects canon_ref only on lane G and lane P (read_boards.py, the if lane in ("G","P")

guard), which is exactly §3.2b's REFUSAL stated as arithmetic — a dispatchable row may never key a

pending: token, and a lane-R / generation_barred row may, because it dispatches nothing. The

guard §3.2c owed was not generation_status but LANE, and it was already armed. Nothing in the

code changes; the sentence now says what the tooth does. The live occasion: six Ch-2 rows were

authored at lane G/P on pending: keys, and R-B re-keyed all six to live primaries rather than

widening the prefix set to admit them.

3.2d THE PLATE ROW COLUMN SET — appended, never reordered

docs/art_boards/PLATES.csv keeps its existing seventeen columns in their existing order (renaming

or reordering them re-baselines every consumer and every fixture) and gains thirteen, appended:

3.2e THE STAGE DISCIPLINE — four candidates are one plate's DRAFT, not four plates

The four-seeds-per-subject sweep is the DRAFT stage of ONE subject. Recording those four as

generation_status=complete makes every downstream count wrong by about 4x and inflates D1 progress

against floors that count FINISHED plates.

• stage carries draft, review or hero, and generation_status then means "complete AT THIS

STAGE" — which is the only reading under which a candidate sweep and a delivered plate are not

byte-identical from the outside.

• A per-class ledger counts each stage separately and joins the counts to density_model.csv's

floors, so "how far is D1-01 against 180" becomes a query. That ledger does not exist yet; it is

named as NOT BUILT here and its rider (FR-084) carries the honest-NONE tooth until it does.

• The floors in density_model.csv are hero-stage counts. A schedule that counts drafts against

them is not optimistic, it is wrong.

3.2f render_tier IS ITS OWN AXIS — RATIFIED

RATIFIED 2026-08-03 (thirty-first sitting, the pipeline-hardwiring mandate). render_tier joins the

CLOSED generation_prompt_hash payload key set as its own key, with the closed vocabulary

ITERATION · REVIEW · HERO_FAST · HERO — the render presets §4 measures. Adding a key to a

closed set is a RATIFICATION and never an edit, which is why it is recorded here with its sitting.

• generation_tier stays what DR-2 §6.1 froze it as: the FIDELITY and BUDGET class, closed at

background · region_standard · region_hero · realm · apex. It is not a render preset and

it never was.

• The 2026-08-02 lane wrote generation_tier: "REVIEW" into 76 stored hashes — the render preset's

name inside the budget-class enum. That is the exact five-senses-of-tier collision

docs/ASSET_DROPIN_CONTRACT.md §1 bans a bare tier column to prevent, and it happened one level

down where no tooth was looking. Two axes, two keys, two closed vocabularies is the fix.

• Because render_tier folds into the hash, a plate re-rendered at HERO gets a different filename

from its REVIEW predecessor by construction — the same self-enforcement the licence class already

gets, applied to resolution.

3.2g The factory riders this section arms

docs/FACTORY_CONTRACT.md carries the per-chapter obligations that ride this section, so the art

program is measured by the same machinery that measures everything else: FR-083 (key every plate

on a live canon id), FR-084 (deliver against the per-class floors, counted not asserted),

FR-085 (the Lane-G care refusal, as arithmetic), FR-086 (a resolvable licence record per

generated plate), FR-087 (previz honesty and the marketing-clear publish gate). All five are

SEED: they surface in every chapter checklist and gate nothing yet, which is the honest state.

3.3 What git carries, and the arithmetic that forbids the alternative

Git carries rows and hashes only — the landed rule (docs/art_boards/README.md: "Plate BINARIES

never enter this repo"). The arithmetic behind it, now that the corpus is sized:

Storage does not constrain this program. Git would have — by four orders of magnitude. The

pointer discipline is not fastidiousness; it is the only thing that keeps the repo clonable.

3.4 How the gate verifies a pointer resolves — LANDED: harness/check_art_pointers.py

Specifying this was the brief; a landed gate beats a spec, so it is landed and mutation-proven.

The hole it closes. read_boards.py T-PLATE-FILE checks that a declared asset_path exists and

is non-empty (read_boards.py:33). sha256 is a live PLATES.csv column carried into the critics'

worklist and no tooth in the 42-gate roster reads it. Today, before this commit: a plate file

swapped on disk passes; a complete row with an empty sha256 passes; a regeneration that

overwrites its predecessor passes; and the entire table passes on any machine where

HUMANITY_ART_ROOT is not mounted, because a path that is never resolved never fails.

P7 is the tooth this repo's own history demands: GATE 36 and GATE 39 both exist because a check

that cannot match reads as a check that passed. An unmounted volume is that same defect wearing a

drive letter.

Proof, not assertion. 8 teeth · 20 must-fire fixtures + 2 must-not-fire baselines, run on every

real invocation (exit 3 if a ruler rotted). All 8 were additionally mutation-proven **live against

docs/art_boards/PLATES.csv**, each mutation firing its named tooth and no other, with the file

restored byte-exact (sha ed9dd9ec… before and after). One mutation — complete with an empty sha,

on a machine with no art root — legitimately trips two teeth (P1 and P7); re-run with a root

present it fires P1 alone, which is the isolation proof.

It runs green today: MODE=MANIFEST_ONLY declared=0 complete=0 verified=0 PASS. Every plate in the

repo is still pending, so there is nothing yet to verify — and the gate says which of those two

things it means, which is the point.

---

4. THROUGHPUT — MEASURED ON THIS BOX, 2026-08-02

Not a published benchmark. A real timing sample against the live ART ComfyUI instance

(127.0.0.1:8189, the only instance that exists on this box —

docs/5090_SETUP_RUNBOOK.md:1276), day window, GPU otherwise idle, 16 generations.

Generator: Z-Image-Turbo bf16 (Tongyi-MAI/Z-Image-Turbo, Apache-2.0, SHIPPABLE per

PIPE_ART_2026-07-29.md §1) + qwen_3_4b text encoder, res_multistep/simple, cfg 1.0, shift 3.0.

Record: docs/concept_art/art_throughput_measurements.json. Re-measure with

python harness/art_throughput.py --measure.

Finding worth more than the headline: BATCHING IS NOT A LEVER. Batch 4 is *slower* per plate

(2.91 s) than batch 1 (2.78 s). The queue is compute-bound, not overhead-bound, so no scheduling

cleverness buys throughput — only resolution and step count do. Recorded because the obvious

optimisation is the wrong one and someone will otherwise spend a day on it.

4.1 The four-stage pipeline and the wall clock

A one-stage model is what makes art schedules lie. The real pipeline: DRAFT k=4 candidate seeds

per plate at ITERATION tier (you do not ship the first sample) → REVIEW the one selected

candidate at REVIEW tier for the leaf page → HERO only APPROVED plates at HERO tier → REGEN

10 % of plates repeat DRAFT+REVIEW after a comment. k and the 10 % are PARAMETERS, printed with the

plan so nobody mistakes them for measurements.

Emitted by python harness/art_throughput.py → docs/concept_art/art_throughput_plan.json. Not one

of these numbers is typed into this document by hand.

**The headline: the entire Prologue→Ch 13 art slice is ~6 nights of unattended GPU. The entire game

is ~33 nights.** Against ~8.8 h and ~52 h of Josh respectively. The GPU is not the constraint and

never was — the review surface is, which is what §2 is for.

4.2 Night window versus day window — the VRAM law, now measured

docs/5090_SETUP_RUNBOOK.md:1572 (D-11b) rules: *"keep 04:00 as the serialized nightly reservation;

never overlap asset-gen on VRAM."* That ruling is now also a measurement: the ART lane peaks at

20,653 MiB of 32,607, leaving 11,954 MiB — not enough to hold a UE 5.8 editor session with a

loaded landscape beside it with any margin.

• NIGHT (20:00→04:00, 8 h, GPU exclusive, unattended): the DRAFT and HERO queues. Drains before

the 04:00 soak and does not resume until the soak reports.

• DAY (attended, contended): REVIEW-tier single-plate regeneration only — 5.35 s, one plate, on

demand while Josh is actually commenting. A comment answered in the same sitting is worth more

than a night of throughput, and it fits in the headroom.

• NEVER: HERO renders in the day window. A hero plate is spent *after* a human said yes; it can

always wait for the night.

4.3 The honest caveat: this is not the ruled primary

PIPE_ART_2026-07-29.md §1 rules HiDream-O1-Image-Dev (MIT, 28 steps) the T2I PRIMARY and

Qwen-Image-2512 the co-primary. Neither is installed —

models/diffusion_models/ holds z_image_turbo_bf16 only. Every figure above is therefore a

ceiling on speed and a floor on cost.

Isolating the fixed term from the two 2048² points — (34.38 − 15.04)/12 = 1.612 s/step, fixed

= 2.147 s — and scaling the sampling term to 28 steps: **the full-game target becomes ~396 GPU-hours

(≈49.5 nights)** if HiDream replaces Z-Image at equal resolution. The schedule is stated at both ends

so nothing rests on the fast model staying the pick. (The ITERATION multiplier ×1.57 uses the

2048²-derived fixed term, which understates it at 1024²; treat it as a lower bound.)

---

5. THE APPROVAL LOOP

Josh's ask: *"you develop those pages and I approve them or make comments on how to adjust or

modify."* Four mechanisms, all four already exist — this section wires them, it mints nothing.

5.1 The page's response surface

The leaf is published as a default-private artifact page (the ruled medium —

docs/art_boards/README.md "The board medium is RULED"), re-published to the SAME url recorded in

MANIFEST.csv.published_url. Under the rows it renders a PICK BLOCK: pre-formed lines, already

in the ruled intake grammar, behind one copy control.

- [ ] BRD_0142 — A2 · B1 · C-SQ-3
- [ ] PLT_0142_CR_0001_sil — regen: too many dorsal spines, the real Komodo reads smoother
- [ ] PLT_0142_CR_0010_beh — hold

That grammar is not invented here. It is docs/JOSH_REVIEW_INBOX.md's ruled line format

- [ ] <TARGET> — <comment>, and python harness/queue.py intake already turns each line into a

work-queue item with its gates, its critic and its derivation attached, ticking the line to

- [x] WQ_00NN in place so intake is idempotent. **Josh copies one block on his phone and pastes it;

nothing is transcribed by hand and nothing is dropped** — an unrecognised target becomes a BLOCKED

item with reason unresolved_target, never a silent drop.

Dependency, stated plainly: docs/JOSH_REVIEW_INBOX.md, harness/queue.py and

docs/AUTONOMOUS_RUN_LOOP.md are the run-loop lane's artifacts, landed at 04aa2a7e

("GATE 42 queue"). Its BRD_ target row already routes a board id to a concept-art-region

regeneration item with gates boards + canon_derivation and the four image critics attached —

which is exactly the receiver §5.2 needs, built independently and arrived at from the other side.

What that row does not yet resolve is a PLATE id. A leaf comments per row, not per board, so

PLT_… must join queue.py's target resolver or every plate-level comment lands as a BLOCKED

unresolved_target item. That is one resolver case, it is named here rather than assumed, and

until it lands the honest fallback is a board-level comment naming the plate in its text.

5.2 How a comment becomes a regeneration — using the landed states, not new ones

1. intake creates the work item. The comment text is preserved verbatim.

2. The comment joins the prompt payload. harness/generation_prompt_hash.py recomputes

generation_prompt_hash over the closed key set. The hash necessarily changes.

3. A changed hash is the landed regeneration trigger (docs/generation_lifecycle.json

trigger_column: regeneration_trigger): the row moves to regeneration_required — one of the

five ratified states, not a sixth — and regeneration_count increments.

4. Because the hash changed, <hash12> changed, so the new plate writes to a new path. The

predecessor survives as the evidence a rejected output is required to be. GATE 43 P3 is what

makes that true rather than intended.

5. first_pass_rate = boards_with_zero_regenerations / boards_reviewed stays a query over

MANIFEST.csv — which is why regeneration_count is a column and not a memory

(docs/art_boards/README.md "The review cadence is RULED"). It drives the ruled cadence: one leaf

at a time until ~90 % first-pass over a trailing ten, then all remaining at once.

Nothing in that chain required a new state, a new column, or a new hash function. The licence

self-enforcement rides along free: a model demoted to previz_only changes the same hash and fires

the same regeneration, with no one re-reading a licence.

5.3 What an approval becomes

An approved plate is not canon and does not become canon. It becomes a **locked

canon-adjacent reference** by being written into the art bible — docs/art_bibles/<subject>.md, which

REALM_DESIGN_PROGRAM.md §3.6 makes the only artifact the asset factory reads, precisely so

boards never become a second source of truth. The bible field carries plate_id + sha256 +

generation_prompt_hash.

Three properties follow, all mechanical:

• Downstream cites the bible, never the board. assetgen-3d (docs/task_router.json) reads the

bible; the bible names the plate; GATE 43 P6 proves the plate's bytes are the approved bytes. The

chain from a 3D asset back to the image Josh approved is hash-verified end to end.

• A swapped file breaks the chain loudly. Change the PNG on disk and P6 goes red. "Locked"

means locked to a sha, not to a filename.

• Ratification stays the director's. docs/task_router.json › art-bible: a subagent may create

AWAITING PICK; only the director writes RATIFIED, "a subagent writing a user ruling into canon

reads as fabricated authority." This program does not touch that line.

---

6. PRIORITY ORDER FOR THE PROLOGUE→CH 13 SLICE

Ranked by what unblocks the most downstream work per plate, not by cheapness. The tier column of

density_model.csv supplies P0/P1/P2; the ordering *within* and the two hard predecessors are this

document's, and are named as such.

RANK 0 — THE HARD PREDECESSOR THAT IS NOT ART

Populate the rows the art is for. T0_Equipment_Registry holds 1 row;

T0_Environment_Grammar_Registry 2; T0_Quest_Seed_Registry and T0_RNG_Drop_Table 0; and

the density model budgets creatures against 420 where 150 exist. You cannot generate a plate

for a row that has not been authored, and a leaf cannot enumerate rows that are not there. This is

registry work (docs/task_router.json), it is zero GPU, and it gates everything below it.

Naming it Rank 0 is this document's call, and it is the single most important line in this section.

RANK 1 — THE LAWS (P0, ~465 plates, 6.8 GPU-hours — under one night)

One pick sets a rule that applies hundreds of times. Every one of these built *after* dependent art

means regenerating that art.

RANK 2 — THE REGION LOOK, in build order

D1-01 region key art (180 slice plates) per leg, in the ruled build order Ch 2 → Ch 13. Key art

sets palette and light; §7's palette-conformance tooth measures every later plate in that region

against it, so the key art must be picked before the region's volume runs.

RANK 3 — THE PER-LEG CONTENT, riding each chapter leg

Anchor leaves in node order, each derived from that node's ## Asset anchors: creatures and fauna →

folklore beings → NPC personas → bosses → flora/minerals/gather → festivals → interiors. Ch 2 first,

because CH_02.md's anchor block is already complete, already id'd, and is the node the 2026-08-01

failure was supposed to have realized.

RANK 4 — SLICE-END

D1-32 realms and the fairy-realm material for the Prologue and Ch 1. This is not a scheduling

preference; it is REALM_DESIGN_PROGRAM.md's ruling that the fairy realm and Prologue/Ch 1 are built

LAST, on the matured concept-art pipeline — restated in CLAUDE.md § Model seats.

Deferred out of the slice

D1-17 legendary species, D1-24 regional weapon skins (no formula possible — the multiplier is

unruled), D1-25 legendary bespoke weapons, D1-36 cinematic key frames, D1-38 schools, D1-39 runes,

D1-41 dungeon interiors. density_model.csv sizes each at 0 for the slice; that is the model's

own arithmetic, not an omission here.

---

7. THE ANTI-SLOP BAR AT VOLUME

Four adversarial critics per board is the landed bar (docs/art_boards/README.md "The four critics

run BEFORE Josh sees anything"). It does not survive 17,876 plates — and pretending otherwise is

how a quality bar becomes ceremonial. The bar is therefore split three ways by what each mechanism is

actually good at.

7.1 DETERMINISTIC — 100 % of plates, zero model cost

Already landed and running on every plate: read_boards.py (18 teeth),

check_canon_derivation.py (10 teeth, D2 residue at zero), check_art_pointers.py (8 teeth).

PROPOSED additions — image-space measurements, numpy/PIL only, ~ms per plate, no model. These are

not taste; each is an arithmetic property whose violation is always a defect:

7.2 SAMPLED — the four critics, as a tripwire for SYSTEMATIC defects

Generation defects at volume are systematic, not independent: a bad prompt template, a wrong

LoRA, a mis-set palette. Systematic defects are exactly what a small sample catches.

To catch a defect running at rate *d* with 95 % confidence, sample *n = ln(0.05)/ln(1−d)*:

The rule: 14 plates sampled per class × leg. Any systematic defect running at 20 % or worse is

caught with 95 % confidence. And the response is not per-plate: **one sampled failure rejects the

whole batch** back to regeneration. The sample is a tripwire, not a verdict on the plates it drew.

That is a ~140× reduction in critic invocations against per-plate review, with a stated and

defensible confidence — instead of an unstated one.

7.3 EXHAUSTIVE — never sampled, no matter the volume

Every care-gated subject takes all four critics, every plate. The scale is known: **47 of 88

slice site rows are care-gated (53 %), and independently 46 of 84 zone rows are protected true**

with 51 of 84 at care_tier elevated (CONCEPT_ART_PROGRAM.md §0.1, whose zone figure this

document re-counted from the region pages and confirms at 84).

The principle: sample the taste questions, exhaust the care questions. A cultural-authenticity

defect on a protected living tradition is not a statistical matter and a 95 %-confidence tripwire is

not an acceptable answer to it. This also honours the CVD §17.1 calibration in BOTH directions —

docs/art_boards/README.md is explicit that care-tier inflation is itself a defect and "a critic that

only ever pushes toward restraint produces reverent empty places."

7.4 HUMAN — what Josh still sees

• Every pick. 142 leaves for the slice, ~833 + ~20 for the full game. He never sees a plate that

failed a floor; that is the landed rule and it does not change.

• A forced random spot-check from the parametric tier. The 2,851 auto-approved parametric units

(§1.1) are covered by a law he picked, but a law can be picked correctly and applied wrongly. Each

batch surfaces 3 randomly drawn parametric plates on the leaf that set their law. It costs seconds

and it is the only thing standing between "the ladder was approved" and "the ladder was rendered."

• Nothing else. Not the drafts, not the rejects, not the sampled critic passes.

---

8. WHAT THIS COMMIT LANDED, AND WHAT IT DID NOT

Landed and running:

• harness/check_art_pointers.py — GATE 43 art_pointers, 8 teeth, 20 must-fire fixtures + 2

must-not-fire baselines, all 8 mutation-proven live against PLATES.csv with byte-exact restore.

• harness/art_throughput.py + docs/concept_art/art_throughput_measurements.json +

art_throughput_plan.json — every figure in §4 emitted, re-measurable with --measure, 7 self-test

fixture groups including a monotonicity check on the committed record and an integer-parse check on

every density row (a silent 0 would shrink the schedule and look green).

• This document.

NOT landed, and named so it is not mistaken for done:

• The §7.1 image-space measures are PROPOSED. They are a second gate (art_quality) and a second

commit; nothing in §7.1 is running today.

• The PLT_ target case in harness/queue.py's intake resolver (§5.1). Board-level comments

route today; plate-level ones do not.

• The leaf page renderer. harness/boards/render_board.py renders the landed 3-panel board; the

12-row leaf is a new template.

• No plate has been generated for canon. The 16 images produced for §4 are throughput instrumentation,

live under D:\ComfyUI\output\timing_*, and are not canon-adjacent, not registered, and not

routed to the art root.

---

9. PROVENANCE — how every number here is re-derived

The two numbers in this document that are NOT measured are §2.2's 12 s / 45 s review times and its

20 % comment rate. They are labelled MODELED where they appear, the sensitivity is tabulated at caps

10/12/16, and they should be replaced by a measurement after the first ten leaves — at which

point §2.2's table is re-run and the cap is re-derived rather than defended.

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