comic-style-bible-lock — the Style Bible as an Executable Gate Rubric (Phase 1 · S2)

The S2 step of [`comic-author`](../comic-author/SKILL.md): turn ART_BIBLE.md from a wall of taste adjectives into a machine-addressable convergence target. The bible is the single visual dialect that is (a) read verbatim into every bake prompt so each frame converges to the same look, and (b) read verbatim into every visual reviewer's prompt so the cross-model panel_gate scores style_consistency / identity_consistency against a ground-truth spec — not against the reviewer's vibe. The bible literally declares this of itself:

style_consistency / identity_consistency 的字面定义就是"符合本文件",不是凭感觉。 无 ground-truth spec,跨模型 gate 会永远吐"style drift"却没有收敛方向。 — examples/comic_m3_audit/ART_BIBLE.md, lines 3–5

That second line is the whole reason this skill exists. Without one source of truth a cross-model gate emits endless "style drift" with no convergence direction, and the film silently grows two visual dialects. This skill locks each dimension to a named anchor so the gate has an individually-addressable, source-pinned target instead of one prose blob — one visual dialect, never two.

  ART_BIBLE.md (taste, user-approved, dated, probe-anchored)
        │
        ▼
  ① bootstrap  ── deterministic parse → ONE immutable style_anchor node per dimension
        │           (palette · line · shading · two-world warm/cyber lighting · identity-lock
        │            · per-world STYLE_PREFIX · text-mode · forbidden · by-design exceptions)
        │           +  EXACT style_bible gate (6 dims ≥4, 6 vetoes)
        ▼
  ② validate <authoring_text>  ── deterministic lint (NO LLM): FORBIDDEN vocab + identity-conflict + world-mismatch
        │                          → called by every downstream author step BEFORE the text enters a bake prompt
        ▼
  ③ update  ── (after ≥1 KEEP panel) Gemini-harvest recurring drift → APPEND-ONLY by-design exception anchors
                (≥50% same-beat quorum, cross-model only) — never rewrites the baseline

Role

Compile, lock, and police the project's locked visual identity across the comic lifecycle. You own the 3-subcommand lifecycle — bootstrap (bible → immutable anchors + gate), validate (deterministic lint of candidate authoring text), update (cross-model drift-harvest into append-only exceptions) — and the EXACT `style_bible` gate below. You do not invent taste: every anchor traces to a user-approved, dated, probe-anchored line in ART_BIBLE.md. If the bible is missing you auto-bootstrap a scaffold and escalate to the user for approval; if it exists and is locked you leave it locked and only ever add by-design exceptions. The convergence guarantee — "one visual dialect, never two" — is enforced here or nowhere.

Constants

• BIBLE = <project>/ART_BIBLE.md — the source of truth; version-stamped + dated + project-pinned

(real: version art-bible/1.0 · 锁定 2026-06-08 · 项目 aris_comic_v1).

• ANCHOR_DIR = <project>/wiki/nodes/ — one style:<slug>.json per dimension (node_type: style_anchor).
• DRIFT_REVIEWER = mcp__gemini__analyzeFile (auto-gemini-3) — the visual drift reviewer for update.

Cross-model independence: the executor (Claude) only aggregates Gemini's JSON; it never judges drift itself.

• QUORUM = (n+1)//2 (≥50%) of the same-beat multi-panel drift samples must flag a dimension before an

exception is written (bump to 75% if drift-from-noise floods). One noisy panel must not mutate the bible.

• MAX_PANELS_PER_UPDATE = 6 (cost cap on the Gemini drift pass).
• GATE = the deterministic style_bible gate (6 dims ≥4 + 6 hard vetoes) — **rule-check, NO cross-model

gate* (the bible IS the rubric the other* skills' cross-model gates consume; here we only certify it compiles).

• IMMUTABLE BASELINE — bootstrap anchors and the §0–§7 spine are locked. update only appends

exception anchors; pivoting the baseline = a fresh bootstrap in a NEW project dir, never an in-place edit (the cross-model gate must always compare new output against the ORIGINAL anchors). See acceptance-gate (locked=immutable).

The §0–§7 section spine bootstrap asserts (the worked example's structure)

bootstrap requires the bible to carry this exact spine — each section is the source of one or more anchors. Copy examples/comic_m3_audit/ART_BIBLE.md as the shape:

Procedure

① bootstrap — compile the bible into immutable anchors (+ certify the gate)

Run once per project, before any other S-step authors text. Re-run is idempotent (overwrites byte-identical).

1. Assert the bible exists. If <project>/ART_BIBLE.md is missing → auto-bootstrap: write a §0–§7

scaffold from the table above with every taste field a TODO(user-approve) placeholder, append a log line, and STOP — escalate to the user for the §0 probe-image + §0.5 world choice + §1 hex table + §6 text-mode approval. Never fabricate a probe image, a hex, or a user-approval date. (exit 2)

1. Assert version + lock + project stamp + user-approval dates. Grep for the version art-bible/… line,

the 锁定 <date> lock, the project id, and at least one dated user-attributed approval line ((用户 <date> …)). Any missing → veto (see gate); do not proceed.

1. Assert the §0–§7 spine compiles. Grep each section header from the table above. A missing §0 (no probe +

approval), §0.5 (no by-design instruction), §0.6 (no STYLE_PREFIX[<world>] line for a world some panel uses — build_prompt.py fail-closes, exit 7, on every panel of that world), §1 (no authority ref / hex table), §6 (no text-mode), or §7 (no FORBIDDEN list) is a hard veto.

1. Deterministic parse → ONE `style_anchor` node per dimension. Regex-grab the §0.5 palette block, the

§0.6 STYLE_PREFIX[<world>]: lines (one anchor per world, payload.value = the verbatim ≤200-char prefix), the §1 hex table, the §3/§4 rules, the §6 mode list, the §7 FORBIDDEN list, and the §0.5 by-design exceptions. Emit exactly the anchors in the table's right column. Node shape (matches schemas/node_schema.json → style_anchor, payload required dimension, value, source, locked):

   {
     "node_id": "style:identity_executor",
     "node_type": "style_anchor",
     "title": "identity-lock — executor (blue hoodie, brown hair, NO beard)",
     "status": "locked",
     "created_at": "2026-06-08T00:00:00Z",
     "tags": ["style_anchor", "baseline", "identity"],
     "payload": {
       "dimension": "identity:executor",
       "value": {"hoodie": "#1D4684", "hair": "#925935", "beard": false,
                  "iron_law": "executor 永远蓝+棕发+无须;L/R 位置可变,颜色/胡子/体型轮廓不可变"},
       "source": "ART_BIBLE.md §1 (authority duo_canonical_ref_v001.png)",
       "locked": true
     }
   }

Every anchor sets status: "locked", payload.locked: true, payload.source = the exact §-section it came from, and tags include ["style_anchor","baseline"]. The §6.5 density anchor additionally carries payload.value.hard_reject: false (quality gradient, NOT a hard reject).

1. Run the EXACT `style_bible` gate (below). All 6 dims ≥4 and no veto → write the anchors, append a

line to <project>/wiki/log.md, print {anchors_written, n_anchors, gate}. Any dim <4 or any veto → fail closed, write nothing, print the failing dim/veto + the §-section to fix (exit 1).

1. Wire the convergence target. The bible pointer lives in the **pointer-hub manifest

movie.project.json, field art_bible — NOT in `comic.json` (comic.json has no `style_bible`/`art_bible` key; its top keys are `schema_version/comic_id/asset_root/meta/defaults/ui_tokens/identity_refs/pages/panels`, and [`comic-json-compiler`](../comic-json-compiler/SKILL.md) is the owner that emits `movie.project.json` as "the pure pointer-hub manifest"). Confirm `movie.project.json.art_bible` resolves to this project's `ART_BIBLE.md` (and `movie.project.json.comic_json` to `comic.json`); if the `art_bible` field is missing, escalate to comic-json-compiler to add it — do not invent a `comic.json` edit. Then confirm the per-world prompt seam: the shipped [`comic-panel-prompt-builder`](../comic-panel-prompt-builder/SKILL.md) runner (`build_prompt.py`, `read_style_prefix()`) reads the §0.6 `STYLE_PREFIX[<world>]` line for each panel's `world` DIRECTLY from ART_BIBLE.md (regex over the bible text; ≤200 chars enforced; exit 7 if the world's line is missing) — the `style:style_prefix_<world>` anchors mirror those lines for wiki addressability, but the bible line itself is the runtime truth the builder consumes**. This is the "one dialect into every prompt" guarantee, made mechanical.

② validate <authoring_text> — deterministic lint (NO LLM)

Called by every downstream author step (intent / outline / storyboard / blueprint / prompt) BEFORE the text is allowed into a bake prompt. Pure Python, no model call — fast, reproducible, fork-proof.

1. FORBIDDEN-vocab scan of §7 (whole-word, re.IGNORECASE): any hit on the §7 ban list (voxel 立体块 /

写实照片质感 / non-canonical mascot colors / 非整数缩放 / smooth gradient on a character or solid foreground prop) → passes=false, return forbidden_hits[].

1. Identity-conflict lint against style:identity_*: flag text that puts the reviewer without a beard,

the executor with a beard, or any mascot color outside {blue,green,brown,black} → passes=false.

1. World-mismatch lint against style:by_design_warm_cyber: flag only mis-colored worlds — "a warm

scene gone cold" or "a digital/technical artifact gone warm" — and NEVER flag the warm↔dark split itself (it is by-design). This is the concrete embodiment of intended-variation ≠ drift.

1. (Optional, mostly DROPPED for us) over-prompt / quality-padding guard — the video source banned camera/

lighting/8K/hyperrealistic vocab because it hijacks Pippit's per-beat decomposition and triggers a silent model downgrade. the sidecar image bake has no per-beat decomposition to hijack, so this whole category DROPs by default; keep only a thin quality-padding check if a baked title ever shows the over-prompt signature.

1. Conservative-refuse rule: prefer a false-positive (force a human glance) over letting an

identity/world/forbidden violation slip into a paid bake. Merge → passes_all; exit 0 (clean) / 1 (fail).

③ update — Gemini-harvest drift into append-only by-design exceptions

Run only after ≥1 KEEP panel exists (else exit 2). This folds the bible into the spiral's wiki memory — recurring, intended deviations become written exceptions; one-off noise does not.

1. Gather the most-recent KEEP panel images of a beat, capped at MAX_PANELS_PER_UPDATE, into a manifest.

(Image-panels swap the video source's "frames" for PNG panels — one image per panel, so the quorum is across panels in a beat, not within-clip frames.)

1. Per panel, call `mcp__gemini__analyzeFile` (`auto-gemini-3`) extracting ONLY, in strict JSON, with the

baseline anchor values pasted inline: color_palette (3–5 hex), lighting (dir / hardness / temp), subjects_present, subjects_absent_but_expected, drift_from_bible (item-by-item deviation vs each baseline anchor), safety_flags (faces / logos / celebrity). Gemini, not Claude, is the visual judge — if Gemini-MCP is down → exit 3, NEVER fall back to a Claude-only read (that violates cross-model independence).

1. Aggregate (Claude's only job here). A per-beat exception locks only when a dimension's drift is

reported in ≥ QUORUM panels of the SAME beat. Then: write a NEW style_anchor (tags: ["style_anchor","exception"], payload.value adds support_frames/total_frames, payload.locked: true), and add a uses_style_anchor edge to wiki/edges.jsonl in the exact `{src, dst, type}` shape `cli/validate_wiki.py` enforces (line 158: every edge needs src/dst/type; line 161: type ∈ EDGE_TYPES; lines 165-167: both endpoints must resolve to a wiki node_id or a comic.json panel anchor panels.<id>.wiki_node_id). Endpoints: src = a representative KEEP panel of the beat addressed by its legal panel anchor panel:<panel_id> (e.g. panel:s12_aris_comic_v1, copied from comic.json panels.<id>.wiki_node_id — never a <beat_id>, which is not a legal node_id prefix; the only legal prefixes are intent|style|outline|asset|storyboard|panel|blueprint|prompt|motif|cont|attempt|review|decision|fail), dst = the new style:<anchor_slug> exception node. Carry the evidence on extra (validator-ignored) keys:

   {"src": "panel:s12_aris_comic_v1", "dst": "style:s12_env_density_exception",
    "type": "uses_style_anchor", "evidence": "drift in c/n panels", "reviewer": "Gemini"}

Then APPEND a line to the bible's ## Per-beat exceptions section: - **<beat>** — <dim>: <val> (observed in c/n panels, Gemini).

1. Never rewrite the baseline. update is append-only. Log the append; print {exceptions_added, anchors}.

EXACT gate — style_bible (ported verbatim from the aris_movie source · deterministic, NO cross-model gate)

bootstrap certifies the compiled bible against this gate. It is a deterministic rule-check — there is no cross-model judge here, because the bible is the rubric every other skill's cross-model gate consumes; this gate only proves the rubric itself is complete and conflict-free.

Dimensions — ALL must score `≥4`:

• `anchor_completeness` — every §0–§7 dimension compiled to an addressable style_anchor node (no prose-only blob).
• `authority_ref_resolved` — the §1 identity authority image exists on disk. **Resolve via

movie.project.json.identity_ref (the pointer-hub field) or a basename match under the project** — NOT by the §1 pointer's literal path: in the canonical worked example the §1 pointer (movie-wiki/assets/refs/sprites/duo_canonical_ref_v001.png, echoed by comic.json identity_refs) is a historical private-layout path that does NOT resolve as written; the real file is assets/duo_canonical_ref_v001.png and only movie.project.json.identity_ref points there. Scoring the literal §1 path would false-veto the very bible this skill says to copy.

• `identity_lock_specificity` — per-character hex + beard + silhouette iron-law is concrete, not adjectives

(executor 蓝+棕发+无须 / reviewer 绿+黑发+有须; L/R free; color/beard/silhouette immutable).

• `text_mode_contract` — §6 declares the mode vocabulary (html | baked | code) AND **every observed

text_mode is licensed by an explicit §6 clause* (the checkable rule, NOT a global equality): read `comic.json defaults.text_mode` (and `movie.project.json.text_mode_default`) plus every `panels.<id>.text_mode`, collect the distinct observed modes, and require each one to be permitted by a named §6 clause. (Scores `<4` only if §6 omits a mode that some panel actually uses.) NB §6 is a decision ledger that licenses multiple modes at once — in the worked example it carries both the `默认锁 html` clause AND the `2026-06-09 B08 已切 baked` update clause, so a project that is `defaults.text_mode: html` but 23 baked / 1 html is fully licensed* (do NOT score it down for the html-default-vs-baked-majority spread — both are §6-clauseed).

• `by_design_exception_contract` — §0.5 carries the explicit 冷暖是 by-design 的,不是漂移 instruction so

the consuming gate flags only mis-colored worlds, never the warm↔dark split.

• `forbidden_coverage` — §7 FORBIDDEN list exists and covers the known traps (voxel, character-surface

smooth gradient, non-canonical mascot color, non-integer scaling, identity inversion).

Hard vetoes — ANY one fails the gate outright (no score can rescue it):

1. No user-approval / version / project lock — bible lacks a dated user approval, a art-bible/N.N version,

or a project id.

1. No duo authority ref — §1 has no resolvable identity authority image (resolution =

movie.project.json.identity_ref or basename match, per the authority_ref_resolved dim above — never the §1 pointer's literal path, which in the worked example is a dead historical path).

1. Warm/dark not declared by-design — §0.5 missing the explicit by-design instruction (the gate would then

false-flag the intended split forever).

1. An observed `text_mode` has NO licensing §6 clause — some comic.json panel (or defaults.text_mode /

movie.project.json.text_mode_default) uses a mode that §6 never permits (e.g. a panel bakes text but §6 carries no baked clause, so a baked panel would be gated on the wrong fields — safezone/stray_text instead of baked_text_quality + the expected_literals token-diff). Veto fires only on an un-licensed mode — NOT on a mere html-default-vs-baked-majority spread, which §6's multi-clause ledger legitimately allows.

1. No FORBIDDEN list — §7 absent (the lint has nothing to enforce; "style drift" becomes unfalsifiable).
2. An observed panel `world` has NO `STYLE_PREFIX[<world>]` line — §0.6 misses a world some comic.json

panel actually uses. The shipped build_prompt.py (read_style_prefix()) fail-closes — exit 7 — on every panel of that world, so the bible would strand downstream prompt-building; each prefix line must be ≤200 chars with no camera/lens vocab (the builder enforces the length as a hard problem).

Node it reads / writes (per schemas/node_schema.json)

• Reads: ART_BIBLE.md (source of truth) · the pointer-hub manifest `movie.project.json` (the

art_bible field is the convergence-pointer this skill verifies in bootstrap step 6 — not comic.json, which has no such key) · comic.json (read ONLY for defaults.text_mode + per-panel text_mode, which feed the gate's text_mode_contract dim — comic.json genuinely carries those) · KEEP panel_attempt images (for update).

• Writes: one `style_anchor` node per dimension under wiki/nodes/style:<slug>.json. node_id prefix

is `style:` (per the schema's id pattern). The payload's required fields are exactly dimension, value, source, locked:

• dimension — the locked axis (e.g. palette:dark_world, identity:reviewer, line_shading, forbidden,

by_design:warm_cyber, style_prefix:warm-lab, text_mode, env_density).

• value — the concrete spec (hexes / rule text / the by-design instruction / the mode set / the ban list).
• source — the exact ART_BIBLE.md §N the value traces to (un-sourced anchor = refuse, never fabricate).
• locked — true for baseline anchors (immutable) and exception anchors alike.
• baseline status: "locked", tags ⊇ ["style_anchor","baseline"]; exception anchors add the exception

tag + support_frames/total_frames and an incoming uses_style_anchor edge in wiki/edges.jsonl written {src: "panel:<panel_id>", dst: "style:<anchor_slug>", type: "uses_style_anchor"} (a legal panel anchor → the exception node; src/dst/type + resolvable endpoints, per cli/validate_wiki.py).

Two engine contracts this skill upholds (fail-closed)

The bible is what makes the comic-author engine contracts enforceable, so the gate certifies them at the bible layer:

1. Every panel needs a `condition.content_svg` — the §0–§7 dialect is what every deterministic blueprint is

rendered in; the bible must be lockable so a blueprint has a single style to instantiate (no content_svg: null; the engine rejects it).

1. A baked figure-panel needs `condition.expected_literals` — §6's text_mode_contract dim is a **hard

veto precisely so a `baked` panel's gate fields line up: a baked figure-panel swaps `safezone`/`stray_text` for `baked_text_quality` (≥4) plus a blind token-diff of observed_literals vs the authored expected_literals. If a `comic.json` panel uses a `text_mode` that §6 never licenses** (e.g. a panel bakes text but §6 carries no baked clause), veto #4 fires — the gate would check the wrong fields and a baked panel could ship with no expected_literals. (A §6-licensed mix — its html-default clause AND its baked update clause both present — does NOT fire the veto.)

Worked example

The canonical bible is [`examples/comic_m3_audit/ART_BIBLE.md`](../../examples/comic_m3_audit/ART_BIBLE.md) (version art-bible/1.0, locked 2026-06-08, project aris_comic_v1) — copy its §0–§7 spine verbatim. The concrete pattern to copy:

• The bible declares ITSELF the gate rubric (lines 3–5): `style_consistency / identity_consistency 的字面

定义就是"符合本文件" and the load-bearing warning 无 ground-truth spec,跨模型 gate 会永远吐"style drift" 却没有收敛方向`. Copy this header verbatim into any new bible — it is the contract that makes the downstream cross-model gate converge.

• §0 anchors to a NAMED probe + a dated user approval, never adjectives: `定档在 … (= 批准的

probe/S02_codexgen_a01.png 那个程度). Every later taste decision is a **dated, user-attributed line** (§0.5 用户 2026-06-09 选 B; §6.5 用户 2026-06-09 品味校准, 靶 = cal_S12_a01`).

• §0.5 is the by-design exception archetype — 真实世界 = 暖 vs `数字 ARIS 世界 = 暗黑霓虹极客

(dark_navy_void #0A0E27), then the gate instruction gate 评 style_consistency 时:冷暖是 by-design 的, 不是漂移;只标"该暖的场景冷了 / 该暗的数字物暖了"这种错配. This single line is what stops a naive gate from flagging the intended two-world split round after round (the feedback_gate_identical_scores_judge_broken` lesson: design-aware gates are the fix).

• §1 is the identity-lock archetype — authority duo_canonical_ref_v001.png + a per-character hex table

(executor hoodie #1D4684 / hair #925935 / no beard; reviewer hoodie #30582D / hair #0F0D16 / beard) + the 铁律 executor 永远蓝+棕发+无须;reviewer 永远绿+黑发+有须. 左右位置可变…但颜色/胡子/体型 轮廓不可变. → emit style:identity_executor, style:identity_reviewer, style:identity_authority_ref.

• §3 carries a reviewer-calibrated carve-out (dated 2026-06-08): 背景氛围光是允许的 — wiki-void center

glow / bloom / volumetric light MAY use smooth gradients; the no-gradient rule constrains only character surfaces + solid foreground props. This is a drift-that-became-a-by-design-exception — exactly what update produces, here recorded by hand after a reviewer round.

• §7 FORBIDDEN includes the voxel ban (3D 体素立体块 (本片走 flat 像素)) added **after gemini found a

flat-2D↔voxel drift** — the canonical case of a cross-model finding hardening into the FORBIDDEN list.

• §6 carries the full text-mode decision history including the explicit PENDING duel (`仅当实测 AI 烤 CJK

干净 且 用户明确偏好,才把 text_mode 切到 baked;在用户拍板前,默认锁 html`) — the bible is a decision ledger, and the pending item does NOT unlock the default.

The sibling artifacts that consume these anchors — gen/asset_lib.py (palette pinned to the IR's ui_tokens, "so the film never grows two visual dialects"), gen/check_asset_collisions.py (the single-source gate), and comic.json's per-panel condition.identity_desc / condition.world / condition.expected_literals (the runtime fields the engine reads under each panel's condition) — are the proof that a locked bible turns into a mechanical, fork-proof convergence target.

Hard do / don't (earned lessons)

• DO make the bible the SINGLE convergence target read verbatim into every bake prompt and every

visual reviewer's prompt — "one visual dialect, never two". If two prompts disagree on the look, you have two dialects and the gate can never converge.

• DO anchor §0 to a named probe image + a dated user approval, never adjectives; record every taste

decision as a dated, user-attributed line.

• DO keep baseline anchors immutable — update only appends by-design exceptions; pivoting the style

= a fresh bootstrap in a new project dir, never an in-place edit (the cross-model gate must always compare against the ORIGINAL anchors). See acceptance-gate (locked = immutable).

• DO declare warm↔dark as by-design (§0.5) so the gate flags only mis-colored worlds — intended

variation ≠ drift.

• DO auto-bootstrap a §0–§7 scaffold and escalate to the user when the bible is missing; DON'T

fabricate a probe image, a hex, or an approval date.

• DON'T let the executor (Claude) judge visual drift in update — that's Gemini's job (cross-model

independence); Gemini-down → exit 3, never a Claude-only fallback.

• DON'T let one noisy panel mutate the bible — require the ≥50% same-beat quorum before an exception is

written.

• DON'T port the video source's camera/lighting/8K/hyperrealistic banned-vocab wholesale — codex

image_gen has no per-beat decomposition to hijack, so that category DROPs; keep the FORBIDDEN-list + identity + world lint, which is what bites us.

• DON'T ship a bible whose §6 fails to license a text_mode that some comic.json panel actually uses —

veto #4 fires, and that panel could otherwise be gated on the wrong fields (or ship with no expected_literals). (Conversely, DON'T veto a §6-licensed mix — html-default + baked-update clauses both present — just because the project is baked-majority; that is intended, not a contradiction.)

Protocols (governance contracts this skill honors)

• acceptance-gate — locked = immutable: this gate can DRIVE

(bootstrap, lint, append exceptions) but can't ACQUIT the taste — user approval is the hard gate for §0/§0.5/ §1/§6, and a locked baseline anchor is never edited in place (a style pivot is a fresh project, not a rewrite).

• reviewer-routing — the update drift reviewer is Gemini

auto-gemini-3; never downgrade. (The compile-time style_bible gate is a deterministic rule-check with no model call, by design.)

• reviewer-independence — in update, Gemini gets the panel

image + the baseline anchor values inline, never Claude's interpretation; the executor only aggregates the JSON.

• artifact-integrity — every anchor's payload.source traces to a

§-section of the bible; an un-sourced anchor is a refuse-and-escalate, never a fabricated value.

• review-tracing · output-versioning

— log each bootstrap/update to wiki/log.md; version the bible (art-bible/N.N + lock date) and never silently overwrite a locked one.