comic-outline-creator — the Outline + Motif Contract (Phase 1, Layer 2)

The middle-left of Figure 1: take a locked intent_spec (+ the pre-locked beat/shot skeleton it carries — OR, for an idea-first comic with NO pre-locked skeleton, the 3 lenses first synthesize a beat skeleton from the intent's logline / subjects / constraints / format, then adapt it the same way; the synthesized skeleton is recorded so the deltas-vs-skeleton discipline still applies) and produce a single approved `outline_spec` — the 6-column beat table that drives the storyboard, and the continuity-motif arcs that become the storyboard's checkable contract. This is the comic twin of aris_movie's movie-outline-creator (a Layer-2 authoring-time plan gate before any pixels), specialized for our image-comic route: the deliverable is a plan, never a frame.

Two things are simultaneously true, and the whole skill is built around them: (a) the outline is NOT a free brainstorm — it is an editorial adaptation of a pre-locked skeleton (N beats / M shots, beat boundaries fixed), so every candidate must self-report its deltas vs the skeleton; (b) an unbounded "ARIS capability showcase" silently rots into a product manual unless every capability is bound to a concrete story cost and skill-name tags are banned from the 画面. The 3-lens → Codex-synthesis → density-pass → user-approval loop turns (a) into a faithful, debated outline and the 侧栏讲解铁律 + the gate catch (b) every round.

  intent_spec (logline + tagline + dual-identity + LOCKED skeleton)
        │
        ▼  ① FAN-OUT: 3 parallel named lenses over the SAME skeleton (faithful ∥ comicnative ∥ clarity)
        │     each ends with a self-authored "through-line + deltas vs the skeleton" section
        ▼
  ② SYNTHESIZE — Codex (different model family) reconciles the 3 → OUTLINE v2
        ▼
  ③ DENSITY/DESIGN pass — Codex MAY restructure panel count (e.g. centerpiece 4→5 pages, +mid-act page)
        │     AUTHORS the 侧栏讲解铁律 (anti "product manual" guard)
        ▼
  ④ Emit the 6-column BEAT TABLE + the centerpiece sub-plan + the 5 motif arcs (exact per-shot values)
        ▼
  ⑤ READINESS gate (PRE-user, cross-model, blind to orchestrator prose) — scores ONLY the author-controllable
        │   dims; EXCLUDES user_decision_record + the unresolved-questions veto (those are not yet satisfiable).
        │   Two-step (never cold): write review:* nodes + `reviews` edges → THEN comic-cross-layer-gate <id> --gate outline
        ▼ readiness:advance
  ⑥ USER APPROVAL = HARD GATE — present; record ✅ APPROVED + fill 已定决策 + EMPTY 待确认 block
        │   THEN the FINAL gate (now user_decision_record==5 IS satisfiable) → write outline_spec node (locked)
        │ revise ─▶ ② (bounded: MAX_OUTLINE_REVISIONS=3 → auto-abandon → failure_mode node)
        ▼ approve + signed
  outline_spec (locked) → SOURCE OF TRUTH for the storyboard layer

Constants

• LENS_COUNT = 3, mandatory (faithful, comicnative, clarity) — Tier-1 Workflow fan-out

(fan-out-pattern). Fewer than 3 lenses is a hard veto.

• SYNTHESIZER = mcp__codex__codex, model_reasoning_effort: xhigh — a different model family from

the Claude lens authors. Codex both reconciles the 3 candidates AND runs the density/restructure pass; the outline layer is itself a cross-model artifact (executor ≠ reviewer, reviewer-independence).

• GATE REVIEWERS = mcp__codex__codex (gpt-5.5 xhigh) ‖ Gemini (auto-gemini-3, readability dims) — each

sees ONLY outline_spec.json + intent_spec.json (the skeleton) + the lens files, never Claude's prose. They do not acquit; they each persist a review:<slug> node (payload {target_node_id, reviewer, gate_kind: "outline", review_scores:{...}}) with a reviews edge review:<slug> → <outline_id>, and the SOLE acquittal organ comic-cross-layer-gate --gate outline then fuses those score-nodes (D4: never invoke the gate cold — write the reviews FIRST, then call the gate).

• THE GATE IS RUN TWICE (the deadlock fix): a READINESS pass BEFORE user approval (scores only the

author-controllable dims; the gate's own user_decision_record == 5 floor and the unresolved-questions veto are not yet satisfiable, so the readiness reviewers explicitly omit them), then the FINAL pass AFTER the user signs + the 待确认 block is emptied (now user_decision_record == 5 IS satisfiable). See §EXACT gate.

• MAX_OUTLINE_REVISIONS = 3, then auto-abandon → write a failure_mode node so a re-run must diverge. Its

payload is exactly what validate_wiki.py PAYLOAD_REQUIRED demands for failure_mode: {"layer": "outline", "affected_shot_ids": [<the abandoned skeleton shots, e.g. all S01..SMM>], "active": true} (node_id fail:<slug>, node_type failure_mode, status active). There is NO `encoding_style` field in the schema — carry the "a re-run must diverge" positive-invariant in title or a non-required note, never a fabricated required-looking key; the node must pass cli/validate_wiki.py.

• MAX_DRAFT_ATTEMPTS = 3 — schema-repair loop for the outline_spec node; on each fail, codex-reply

quoting the validator errors verbatim. Never silently coerce a value to pass schema (change the narrative + document it, or surface to the user).

• TARGET_PAGE_CEILING (soft cap, exceed ⇒ Codex must justify in rationale.beat_count_justification,

not a hard stop): mvp=2, demo=6, longform=12 beats per tier (ported from MAX_BEATS_PER_TIER).

• MIN_CAST_PER_BEAT = every beat names ≥1 world/scene + the cast in it (text-panel beats: a

text-panel scene + its world). Ported from aris_movie's MIN_ASSETS_PER_BEAT=2; a structural check JSON-Schema can't express.

• NODE = outline_spec (this skill reads intent_spec, writes one outline_spec). See §Node.
• APPROVAL IS NON-AUTOMATABLE — the loop can DRIVE to a clean gate, but only the user ACQUITS the

outline (acceptance-gate). Nothing proceeds to the storyboard layer until the user signs off, recorded inline.

Input contract — what this skill reads, what it must not invent

The contract boundary is the `intent_spec` node (§Node) — produced upstream by comic-intent-parser, carrying: the logline (a single sentence whose climax may be an editorial inversion — e.g. "the climax is NOT winning — it's that ARIS didn't let itself off the hook"); the tagline; the `dual_identity` constraint ("an emotional story of honest research + an ARIS capability showcase") the whole debate optimizes against; and the `source_skeleton` — the pre-locked beat list (B01…BNN), shot list (S01…SMM), and beat boundaries. The skeleton is LOCKED; this skill ADAPTS it, never re-derives it.

This skill is plan + verify: it never bakes a pixel (image gen is comic-director / Phase 2-3) and never invents a beat boundary or a shot. Adding a page (a density-pass split) is allowed and must be accounted for in the deltas; moving a beat boundary or dropping a shot is a delta that must be justified + surfaced to the user. If intent.status != locked → REFUSE (the upstream hand-off token is locked = approved + frozen; an under_review/draft intent is not ready — re-run comic-intent-parser to its user gate). If the skeleton is missing → REFUSE and ESCALATE (do not fabricate one).

Procedure (an agent can execute this step by step)

0. Load + refuse-guard

Snapshot the intent_spec node. Pull cumulative wiki context: prior LOCKED assets (for reuse-over-invention), and failure_mode nodes with layer="outline" (rejected/superseded outlines — a re-run must diverge from them). Refuse if intent.status != locked (the upstream author-node hand-off token is locked, not active). Checkpoint an OUTLINE_STATE.json (resumable — resumable-runs).

1. FAN-OUT — write 3 parallel candidate LENSES over the SAME locked skeleton

This is Tier-1 Workflow fan-out (fan-out-pattern): three independent passes, each a named lens over the identical skeleton, NOT three free drafts. Write each to its own file (story/outline_candidate_{faithful,comicnative,clarity}.md). The three lenses are fixed:

• `faithful` — one-to-one fidelity: M shots → M panels, no cut / merge / add, beat boundaries

B01…BNN preserved exactly. The conservative anchor (palette grammar + stamp mirrors live here).

• `comicnative` — comic-as-a-medium: collapse runtime / motion shots into static-friendly pages

(in our run S07-S09 → one 2×2 page; 22 shots → 12 panels); choose a full-bleed / 2-up / 2×2 page-rhythm for scroll-reading.

• `clarity` — legibility for a non-ML reader: foreground the through-line so the headline number-arc

(in our run +6.2 → +1.4) is readable via one concrete physical prop (the Tok|yo broken-JSON evidence) + one wordless emotional gauge (executor bounce height).

Each candidate MUST end with a self-authored section: `## through-line + deltas vs the skeleton` stating exactly what it merged / added / kept and why (e.g. "merged S07-S09 → one 2×2 page (motion is redundant in a static page); kept all 13 beats; moved Tok|yo's first plant earlier to B06-S09 as a re-pointable物证"). A candidate with no deltas section is a hard veto — delta-accounting is the discipline that proves it is an adaptation, not a guess.

2. SYNTHESIZE — Codex reconciles the 3 → OUTLINE v2

Call the SYNTHESIZER (different model family). Feed it the three lens files + the intent_spec + schemas/node_schema.json's outline_spec shape, inside a demarcated === EXTERNAL CONTEXT (advisory) === fence; Claude's prose lives ONLY in that fence (reviewer-independence). The synthesis takes the best of each lens — in our run: faithful's two-world WARM/DARK palette grammar + the REJECT↔ACCEPT mirror, comicnative's page-rhythm, clarity's motif arcs + countdown spine. Emit {outline, rationale}.

3. DENSITY / DESIGN pass — Codex MAY restructure panel count

Codex then does a density / design pass that may materially change panel count where the read demands it (in our run: the centerpiece B08 went 4 panels → 5 single-read pages by adding an "audit entry" panel, with the 2×2 kept only as recap/hero/GitHub-screenshot; and a new mid-act page B09.5 "research star-map / wiki" was added after the audit, before the paper — flagged as "the most powerful position"; panel count moved 22 → ~24). This pass also AUTHORS the 侧栏讲解铁律 (step 6). Adding pages is fine; account for every one in the deltas.

4. Emit the 6-column BEAT TABLE

Emit the canonical table — fixed columns, in this order:

| Beat (zh/en) | Panels | 布局/layout | 世界/world | 内容 + 台词 gist | ARIS 特性（绑剧情）|

• Beat — bilingual id + name (**B01 卧室·倒计时**).
• Panels — the shot ids it OWNS (S01 · S07,S08,S09大 · S12-15 + 新增 S12a).
• 布局/layout token ∈ {full-bleed, split, 2-up, 1-panel, 竖排3, 5单张页(默认)/2×2 recap}.
• 世界/world token ∈ {暖/WARM, 暗/DARK, seam, 暗·星空/starfield} — human beats warm, ARIS beats

dark-cyber, explicit seam panels at hand-off / payoff. (The warm/cold split is by-design — the downstream gate must never flag it as drift.)

• 内容 + 台词 gist — panel action + baked-dialogue gist + a one-line side-narration (旁白).
• ARIS 特性（绑剧情） — the bound capability, phrased as a STORY COST, never a feature name

(✅ experiment-bridge（证据是暂定的） · honest re-run（自纠 overclaim） · auto-review-loop（对抗审）; ❌ experiment-audit, novelty-check as bare tags). A bare feature name here is a hard veto.

5. Centerpiece sub-plan (the densest beat)

For the centerpiece beat, emit a sub-table | 页/page | 画面/image | 侧栏（先讲发生了什么，再点能力）| with the verbatim 70/30 composition rule:

每页：技术主画面 70% + 角色情绪 30%（右下角蓝 exec bounce 当体温计）+ 侧栏讲解。

Each row's 侧栏 puts the event-sentence first, then · *trait: <skill>* second (so the side-narration reads as story, then footnotes the capability).

6. Embed VERBATIM the 侧栏讲解铁律 (the anti "product-manual" guard)

This block is load-bearing — copy it verbatim into the outline (it is what stops a capability showcase from rotting into a product manual):

🎙️ 侧栏讲解铁律（防止"能力展示"变产品手册 —— codex） - 每条能力必须绑一个剧情代价：不写"ARIS has experiment-audit"，写"executor 的庆祝被审计打断，所以 ARIS 没撒谎"。 - 侧栏第一句永远讲发生了什么；第二句才点能力名。 - 每个能力都改变角色状态（bounce 高低 / 咖啡冷掉 / 红队介入 / claim 被降级）。 - 画面里不堆 skill 名标签；skill 名只作侧栏小字或 hover/click metadata。

7. Lock the continuity-motif arcs AT OUTLINE TIME (exact per-shot values)

The outline doubles as the continuity contract — the storyboard's motif-state table and the panel/assembly gates will be checked against these arcs, so they must carry exact per-shot values now. Lock the 5 arcs (our run's values, as the worked pattern):

• mug heat: STEAMING(S01) → fading(S03) → COLD(S20) — B01/B12 bookend.
• executor bounce (wordless emotion thermometer): `MAX(S02) → HIGH(S09) → DROP(S11) → ZERO/SILENT(S12–S15,

arc low point) → tempered-return(S16) → smallest-warm(S21)`. (Storyboard will add the uniqueness invariant "S02 是全片唯一 MAX"; no post-fall peak.)

• Tok|yo broken-JSON evidence: `born/first-plant(B06-S09) → smoking-gun(B08) → clean footnote(S17) →

faint constellation star(S22)`.

• DDL countdown spine: 24:00:00 → ~19:00 → 00:42 — strictly monotonically non-increasing across

the whole film (storyboard enforces; never rewinds).

• REJECT(S11) ↔ ACCEPT(S16): deliberate mirror stamps so the centerpiece audit reads as the hinge

between them.

8. Write the 6-column table + sub-plan + 铁律 + motif arcs to story/OUTLINE_DRAFT.md

Add a top-of-file Logline + Tagline + 双重身份 block (mirroring the intent), an empty ## ✅ 已定决策 block, and an ## ⚠️ 待你确认 block listing every open design question (e.g. centerpiece 5-page split OK? new mid-act wiki page OK? 侧栏铁律 OK?).

9. READINESS gate — PRE-user, cross-model, blind (NOT the final gate)

This is the author-controllable gate, run before the user has decided anything. By construction the artifact from step 8 has an empty `已定决策` + a non-empty `⚠️ 待确认` block and no user sign-off — so the FINAL gate's user_decision_record == 5 floor and its "open questions unresolved but advancing" veto are not yet satisfiable. Applying the full conjunctive APPROVE rule here would deadlock a fresh agent (stall, or fabricate a user approval). So the readiness pass scores only the dims the author controls and explicitly EXCLUDES user_decision_record + the unresolved-questions veto (see §EXACT gate → READINESS).

Run it as a two-step gate invocation (D4 — never call the gate cold):

1. Fan out the GATE REVIEWERS (Codex gpt-5.5 xhigh ‖ Gemini auto-gemini-3) on the files only

(outline_spec.json + intent_spec.json + the lens files, inside the advisory fence — no orchestrator prose). Persist EACH as a `review` node: node_id review:<slug> (e.g. review:outline_codex_r1), node_type review, payload required {target_node_id: <outline_id>, reviewer: "codex"|"gemini", gate_kind: "outline", review_scores:{skeleton_fidelity:…, delta_accounting:…, coverage:…, page_rhythm_readability:…, story_cost_binding:…, motif_contract_completeness:…, safety_ip:…}} — and write a reviews edge {src: review:<slug>, dst: <outline_id>, type: "reviews"} (the gate walks reviews edges in; reviewed_by is the wrong direction and the fuser won't find it).

1. THEN invoke comic-cross-layer-gate <outline_id> --gate outline to fuse the score-nodes + adjudicate.

(Its own --gate outline floor — min(coverage, identity_lock_feasibility, scene_lock_feasibility) ≥ 4 AND safety_ip ≥ 4 — runs on whatever dims the reviewers scored; the readiness reviewers feed it the author-controllable subset and intentionally leave user_decision_record unscored, which the gate SKIPs, never 0-substitutes.) On a FAIL verdict, fold the blockers and loop back to step 2 (bounded by MAX_OUTLINE_REVISIONS); on a clean readiness verdict, proceed to step 10.

Then schema-validate the (still-`under_review`) `outline_spec` node payload (cli/validate_wiki.py mirrors node_schema.json); on a schema error, run the MAX_DRAFT_ATTEMPTS repair loop with codex-reply quoting the errors verbatim. Also run the 2 structural checks JSON-Schema can't express: (i) coverage — every intent.dual_identity "must-land" item and every skeleton beat maps to ≥1 table row; (ii) MIN_CAST_PER_BEAT — every beat names ≥1 world + its cast. Do not flip to locked yet — that needs the user sign-off (step 10).

10. USER APPROVAL = HARD GATE → FINAL gate → then lock

Present the readiness-clean outline to the user. Nothing proceeds until they sign off. On approval:

• record inline ✅ APPROVED by user <date> (with their words, e.g. ("都可以")),
• fill the ## ✅ 已定决策（用户 <date>） block — **every resolved open question becomes a recorded decision

node** (our run: B06-S09 → big panel; Tok|yo pre-plant → very small, narration does NOT spell it out; B13 constellation → pure glowing star-points, NO labels),

• empty the ## ⚠️ 待你确认 block (every open question is now in 已定决策).

Now — and only now — run the FINAL gate (D4 two-step again): re-fan the GATE REVIEWERS on the now-complete artifact, persist fresh review:<slug> nodes (this round CAN score user_decision_record, since the sign-off + 已定决策 block + emptied 待确认 block exist) with reviews edges, then call comic-cross-layer-gate <outline_id> --gate outline. The FINAL gate applies the full conjunctive APPROVE rule (§EXACT gate → FINAL) including user_decision_record == 5, now satisfiable. On the gate's approve:

• the gate itself mints the `decision` node (node_id decision:outline_<slug>, payload required

{target_node_id: <outline_id>, verdict: "approve", gate_kind: "outline"}) + the decides edge and flips the `outline_spec` node `status: locked` — that FLIP is the gate's job, not this skill's. (Never write status: locked by hand without the gate's approve verdict node and the inline user sign-off.)

• declare the file "SOURCE OF TRUTH for the storyboard layer."

Then hand off to comic-storyboard-creator (it inherits the locked beats + the motif arcs as its continuity ledger).

EXACT gate — this skill's PRE-gate rubric, fused by comic-cross-layer-gate --gate outline

Authority note (D3): the canonical APPROVE predicate + the status FLIP for the outline gate are owned by comic-cross-layer-gate --gate outline (its floor: min(coverage, identity_lock_feasibility, scene_lock_feasibility) ≥ 4 AND safety_ip ≥ 4, plus its outline pre-check that every referenced asset_id resolves and is status: locked). The dims + verdict rules below are THIS skill's pre-gate routing (step 9 step-1 fan-out): the rubric the GATE REVIEWERS score into their review:<slug> nodes, and the local stop-rule this skill uses to loop/abandon before deferring to the fuser. They are not a competing gate — the cross-layer-gate fuses these score-nodes and makes the final call.

Each GATE REVIEWER scores these dims 0–5 seeing ONLY outline_spec.json + intent_spec.json (the skeleton) + the lens files — no orchestrator prose:

• `skeleton_fidelity` — beats / shots / beat boundaries from the locked skeleton are preserved or every

change is an accounted delta (no silent re-derivation).

• `delta_accounting` — every candidate has a deltas vs the skeleton section; the synthesized outline's

added/merged pages are all listed and justified.

• `coverage` — every intent.dual_identity must-land item + every skeleton beat maps to ≥1 table row

with the right cast in must_show (the comic analog of aris_movie's coverage dim).

• `page_rhythm_readability` — layout tokens give a real scroll-read rhythm; the centerpiece is single-read

paced; non-ML reader can follow the headline arc (readability — Gemini co-scores).

• `story_cost_binding` — every ARIS 特性 cell is phrased as a story cost that changes a character

state, never a bare feature name.

• `motif_contract_completeness` — all motif arcs are locked with exact per-shot values (not "a

recurring mug"); REJECT↔ACCEPT mirror + monotonic DDL declared.

• `safety_ip` — no celebrity likeness / copyrighted character / trademarked logo / off-policy content;

anti-IP must_avoid carried into the style bible (HIGHER = better).

• `user_decision_record` — the ✅ APPROVED stamp + 已定决策 block (resolved questions as decision

nodes) + an EMPTY ⚠️ 待确认 block are all present and consistent. Scored in the FINAL round ONLY — at step 9 it is not yet satisfiable, so the READINESS reviewers leave it null (the fuser SKIPs a null dim, it does not 0-substitute).

READINESS verdict rule (step 9 — PRE-user; EXCLUDES the user dim + the unresolved-questions veto)

The author-controllable stop-rule the readiness fan-out routes on (this is what makes step 9 actually passable before the user has decided anything):

skeleton_fidelity >= 4 AND delta_accounting >= 4 AND coverage >= 4
AND story_cost_binding >= 4 AND motif_contract_completeness >= 4 AND safety_ip >= 4

user_decision_record is NOT evaluated here, and the "open questions unresolved but advancing" veto does NOT apply at readiness (the ⚠️ 待确认 block is expected to be non-empty pre-user). Readiness-clean ⇒ present to the user (step 10).

FINAL verdict rule (step 10 — POST-user; the full conjunctive rule, now satisfiable)

skeleton_fidelity >= 4 AND delta_accounting >= 4 AND coverage >= 4
AND story_cost_binding >= 4 AND motif_contract_completeness >= 4
AND safety_ip >= 4 AND user_decision_record == 5

revise ⇐ any blocking dim < 4 but recoverable (loop to step 2, bounded by MAX_OUTLINE_REVISIONS=3). abandon ⇐ safety_ip < 3 OR three consecutive revise rounds with the same root cause → auto-abandon → write a failure_mode node whose payload is exactly the validator-required shape {"layer": "outline", "affected_shot_ids": [<abandoned skeleton shots>], "active": true} (node_id fail:<slug>, status active; no encoding_style key — that is not a schema field).

Hard veto (any one → not APPROVE, regardless of scores):

• fewer than 3 lenses;
• any candidate with no `deltas vs the skeleton` section;
• any ARIS 特性 cell written as a product-manual feature name instead of a character cost;
• a skill-name in a 画面 main label (skill names allowed ONLY in side-narration small-text / hover

metadata);

• open questions unresolved but advancing (FINAL round only — the ⚠️ 待确认 block must be emptied into

已定决策 before the user can lock; this veto is intentionally OFF at readiness, see above);

• motif arcs not locked at outline time (no exact per-shot values).

The loop can DRIVE to a clean gate, but the gate cannot ACQUIT — only the user can (acceptance-gate). A clean cross-model readiness gate is necessary, the signed user approval is sufficient; both are required, and the FINAL gate is what mints the approve decision + flips outline_spec to locked.

Node — reads intent_spec, writes outline_spec

Per ../../schemas/node_schema.json (schema_version node/comic/3.0).

Reads intent_spec (node_id intent:<slug>) — payload required (verbatim from validate_wiki.py PAYLOAD_REQUIRED): raw_input_refs, user_goal, audience, format, constraints, subjects, source_skeleton, dual_identity, uncertainties, confidence. Refuse unless its `status == locked` (the author-node hand-off token is locked = approved + frozen, NOT active; comic-intent-parser's user gate is what flips it).

Writes one outline_spec (node_id outline:<slug> — matches the schema node_id pattern). Payload required fields — every one of these (verbatim from validate_wiki.py PAYLOAD_REQUIRED for outline_spec) must be present: source_intent_id, title, logline, tagline, design_constraint, target_page_count, narrative_beats, beat_table, motif_arcs, character_asset_ids, scene_asset_ids, prop_asset_ids, global_style_bible, budget:

• source_intent_id — the intent:<slug> it adapts. Also write the provenance edge `{src: outline:<slug>,

dst: intent:<slug>, type: "derived_from"} (the author-layer "this node was authored/derived from that node" verb — derived_from is in validate_wiki.py EDGE_TYPES; generated_from is NOT and would FAIL the validator). **The edge record itself must pass cli/validate_wiki.py`** (both endpoints resolve to real node_ids; the type is in EDGE_TYPES).

• title, logline, tagline, design_constraint — carried from intent (the dual-identity constraint).
• target_page_count — within TARGET_PAGE_CEILING for the tier, or justified in rationale.
• narrative_beats — the beat list with each beat's must_show (cast/world per beat — enforces

MIN_CAST_PER_BEAT).

• beat_table — the 6-column table (the canonical render-source for the storyboard).
• motif_arcs — the 5 arcs with exact per-shot values (the continuity contract).
• character_asset_ids, scene_asset_ids, prop_asset_ids — flat lists holding the **node_ids of prior

LOCKED assets when one matches (reuse over invention); emit no new asset node here** — that is the asset layer's job, where reuse_of lives. (There is no reuse_of field on outline_spec — the asset references are these three flat id-lists; don't carry the movie source's asset_plan[].reuse_of here.)

• global_style_bible — pointer to the locked ART_BIBLE.md / style_anchor nodes (incl. the warm/dark

two-world palette + the anti-IP must_avoid).

• budget — page/asset budget for the downstream layers.

Status lifecycle: draft → under_review → locked — locked flipped by the FINAL comic-cross-layer-gate --gate outline approve (after the inline user sign-off), or rejected / superseded. (active is a RUNTIME status, never this author node's hand-off token.) Fail-closed: never write status: locked without the gate's decision node verdict="approve" and the inline user sign-off; never silently coerce a value to pass schema.

Two engine contracts the outline must SET UP (fail-closed, downstream)

The outline does not author blueprints, but it must make them authorable — so the centerpiece sub-plan and any figure-bearing beat are spec'd so the downstream layers can satisfy these without re-litigating. (The runtime field names below are the `comic.json` panel fields the engine reads — condition.content_svg / condition.expected_literals — the only artifact that uses the condition. prefix; do not conflate with the wiki nodes the downstream layers write, where the field is blueprint.payload.content_svg (top-level) and panel_spec.payload.content_blueprint — there is no content_svg on a panel_spec.)

1. Every panel will need a `condition.content_svg` (a deterministic blueprint — figure or layout; the

engine rejects condition.content_svg: null, never write it null). So every figure-bearing beat in the table must name what the blueprint depicts (the curve, the gauge, the broken-JSON chip), not just "a screen" — the storyboard / blueprint authors instantiate the SVG, but the outline owes them a concrete subject.

1. A `text_mode:"baked"` figure-panel will need `condition.expected_literals` (exact numbers/keys,

ascii-tokenizable, verbatim) — or the bake is refused. So every beat whose number is the point (+6.2, +1.4, 0.60→0.71, REJECT, T-00:42) must surface that exact literal in the table's gist as the to-be-gated token; a beat with no audited number → its panel becomes text_mode: "html" downstream. The motif arcs (DDL, exact_parse, claim_delta) are precisely the literal source the gate will token-diff later.

Worked example

A converged, user-approved outline ships in examples/comic_m3_audit/story/ — the ARIS-Movie comic "我把那 24 小时交出去了 / I Handed Over Those 24 Hours". The exact files to copy the shape from:

• The 3 lenses —

outline_candidate_faithful.md (22 shots → 22 panels, beat boundaries preserved, ends with "Through-line + deltas: one-to-one fidelity, …"), outline_candidate_comicnative.md (S07-S09 → one 2×2 page P06; 22 → 12 panels; full-bleed / 2-up / 2×2 rhythm), and outline_candidate_clarity.md (Tok|yo prop + bounce gauge carry the +6.2→+1.4 story for a non-ML reader). Copy the pattern: three named lenses over the same skeleton, each closing with an explicit deltas vs the skeleton section.

• The approved outline —

OUTLINE_DRAFT.md. Copy its exact shape:

• header provenance line: *"OUTLINE v2 — 3-lens debate → codex synthesis → codex design pass (B08 density +

wiki) → user decisions. … ✅ APPROVED by user 2026-06-10 ('都可以') … Now the SOURCE OF TRUTH for the storyboard layer."* — the whole lifecycle is recorded in one line;

• the ## ✅ 已定决策（用户 2026-06-10） block (3 resolved open questions as decision nodes: S09 big /

Tok|yo tiny, narration doesn't spell it out / B13 unlabeled glowing star-points);

• the verbatim ## 🎙️ 侧栏讲解铁律 block;
• the 13-row 6-column beat table (every ARIS 特性 cell is a story cost, e.g.

experiment-bridge（证据是暂定的）);

• the B08 5-page sub-plan with the 70/30 rule + event-first / trait-second 侧栏;
• the ## vs v1 的增量 deltas (22 → ~24 panels, +S12a +S16b) and the ## ⚠️ 待你确认 block.

Copy its shape; swap in your own logline / skeleton / cast.

Hard do / don't (earned lessons)

• DO treat the skeleton as locked and make every candidate self-report its deltas vs the skeleton —

adapt, never re-derive; a candidate with no deltas section is a veto.

• DO write 3 lenses (faithful ∥ comicnative ∥ clarity) and let Codex (a different model family)

synthesize + run the density pass — the outline layer is a cross-model artifact, not a solo draft.

• DO bind every capability to a story cost and embed the 侧栏讲解铁律 verbatim — this is the

guard that stops a capability showcase from rotting into a product manual.

• DO lock the motif arcs with exact per-shot values at outline time — the outline is the storyboard's

continuity contract; vague motifs make the downstream gate uncheckable.

• DON'T put a skill name in a 画面 main label — skill names live ONLY in side-narration small-text /

hover metadata.

• DON'T advance to the storyboard with open questions unresolved — empty the ⚠️ 待确认 block into

已定决策 first; resolved questions are recorded decision nodes.

• DON'T batch ahead, and never draft (or proceed) without explicit user story approval — the user

is the only acquitter; record ✅ APPROVED <date> inline before locking. (Memory: never draft outline/storyboard without explicit user story approval; confirm intent → outline → storyboard layer by layer.)

• DON'T bake or invent — this skill produces a plan; pixels are comic-director's job, and a missing

beat/shot is an ESCALATE, not a fabrication.

• DON'T silently coerce a value to pass schema or to pass the gate — change the narrative + document it,

or surface to the user.

Protocols (governance contracts this skill honors)

• reviewer-independence — the Codex synthesizer + the gate

reviewer get the lens files + intent_spec + schema inside an === EXTERNAL CONTEXT (advisory) === fence, never the author's interpretation; the Claude lens authors ≠ the Codex synthesizer/judge.

• acceptance-gate — the loop can DRIVE to a clean cross-model gate but

cannot ACQUIT; USER approval is the hard gate for the outline (never proceed without it), recorded inline as ✅ APPROVED <date> + the 已定决策 decision block.

• fan-out-pattern — the 3 lenses are Tier-1 Workflow fan-out

(faithful ∥ comicnative ∥ clarity), all reconciled at the single Codex synthesis seat; <3 lenses is a veto.

• reviewer-routing — Codex gpt-5.5 xhigh (synthesizer + gate);

Gemini auto-gemini-3 (optional readability co-reviewer); never downgrade the tier.

• review-tracing — every gate round's reviewer verdict + the user

decision are logged as review / decision nodes so each verdict is auditable.

• output-versioning · resumable-runs

— version the outline (_v{N} + supersedes); OUTLINE_STATE.json makes the loop resumable.