The ref gate the comic side was missing. comic-director's panel_gate reviews composed output; the\nassembly_gate reviews cross-panel consistency. Neither gates the identity refs that feed both — so\nwithout this skill, identity drift is only caught downstream, after credits are spent on a panel bake. This\nskill is the Layer-2 gate on a SINGLE reusable reference asset (the character sheets / location plates /\nprop cutouts the ART_BIBLE.md points every bake at), run before any panel composition.
A reusable asset is a one-way door: once a hundred panels condition on duo_canonical_ref_v001.png, you cannot\nquietly swap it. So the lock must be earned cross-model, never self-certified by the generator. This skill\nlocks an asset only at 准×3 — Claude and Gemini and Codex each independently pass the same review\nround — or it sends the asset back to be regenerated, or (at the round ceiling) escalates the asset\nrequirement itself back to the outline/storyboard layer. The executor (and the generation model family)\nnever acquit their own ref.
Pipeline position (the Phase-1 asset DAG — the documented contract): this lock runs AFTER the\nhuman-approved outline (OUTLINE_DRAFT_VALID only requires every referenced asset to be DECLARED with a\ncomplete, generatable request — never pre-locked assets) and the provisional storyboard's\nconsolidated_asset_requests, and BEFORE OUTLINE_FINAL_LOCK (cheap re-check that the locked assets still\nmatch the approved outline) → storyboard FINAL asset-resolution validation → blueprint authoring, where the\nhard locked-asset barrier sits. (The old single-stage contract — outline gate demanding locked assets —\ndeadlocked and is retired.)
asset (review_status:\"pending\") ◀── comic-asset-ref-generator (S4: generates, NEVER self-locks)\n │\n ▼\n LAYER 1 — STATIC CI GATE (deterministic, zero credits)\n check_asset_collisions.py (one canonical owner per filename — \"one visual dialect, never two\")\n + build asserts: zero-text contract · literal contract · raster-ref 6-field completeness\n │ pass\n ▼\n LAYER 2 — VISUAL LOOP (round R = 1 … MAX_REVIEW_ROUNDS=3)\n ① Claude narrative ‖ ② Gemini visual (primary: identity/iso) ‖ ③ Codex synth (policy/semantic)\n each reads the REAL asset file · scores 5 dims (0–5 ints) · flags failure_modes from the fixed vocab\n │\n ▼ Codex synthesises: MIN-fuse per dim (max for inverted) · threshold · per-reviewer unanimity\n 准×3 ? ── no, a family HARD-FAILS a floor ─▶ comic-asset-ref-generator --regenerate (bump version) ─▶ ① R+1\n │ no, third family just ABSENT (e.g. Gemini down) ─▶ re-VOTE (re-collect the missing reviewer; NO re-bake)\n │ no (safety_ip<3) ─▶ HARD abandon_shot (IP veto, no retries)\n │ no (R == MAX_REVIEW_ROUNDS) ─▶ ESCALATE the asset REQUIREMENT to outline/storyboard (abandon_shot)\n │ yes\n ▼\n LOCK — asset.status=locked + payload.review_status=locked (ONE-WAY DOOR, immutable for the rest of Phase 1)\n + score_progression audit table (ASSET_REVIEW.md) + decision/review/failure_mode wiki nodes + edges\nmcp__gemini__analyzeFile (model: \"auto-gemini-3\", NEVER\nraw gemini-3.1-pro — 429 silent-downgrade to 2.5) ‖ mcp__codex__codex r1 / mcp__codex__codex-reply r2+\n(model_reasoning_effort: \"xhigh\", save the threadId; pin no model — Codex reviewer calls follow the\nlocal codex config, currently gpt-5.6-sol, exactly like the shipped Codex CLI reviewers in run_comic.py,\nwhich pass no -m; only the upstream BAKE payload pins gpt-5.5 + xhigh, the run_comic.get_bake_plan()\ncompat default — a config-driven override there is planned, not yet implemented).approve is NOT enough: cc_approves AND gemini_approves AND codex_approves must ALL be true in the same round (each reviewer's own dim scores\nmust independently clear threshold). Codex-approve-without-unanimity is DOWNGRADED to revise (the \"准×3\nbite\"). prior_lock_count = the count of approvals accumulated this round; lock iff prior_lock_count >= 3.regenerate, the loop STOPS at 3 and\nroutes back to outline/storyboard — the requirement changes; the gate does not grind an image forever.approve can diagnose/veto but is never the sole acquitter;\nthe lock needs Gemini approve + Claude approve too (acceptance-gate.md: the loop DRIVES, it can't ACQUIT).safety_ip < 3 from ANY reviewer = immediate HARD abandon_shot, no\nretries (generalises comic-director's anatomy single-vote veto to IP/likeness).--lite is NON-ACQUITTING — when Gemini is unavailable it may set only a PROVISIONAL\nstatus: under_review (tag lite_provisional: true), NEVER a root status: locked. The one-way LOCK\nrequires the THIRD family (Gemini) to lock-pass in a later round: 准×3 is CC AND Gemini AND Codex same-round\nunanimity (canon), and the only sanctioned relaxation is the gate's \"a model is UNAVAILABLE → threshold-only\nPROVISIONAL, exit 2\" — never a Gemini-absent lock. Use --lite only to keep progress while Gemini is down;\nnever the default; it can never write the locked one-way door.examples/<project>/story/ASSET_REVIEW.md (the human score-progression audit) + wiki nodes/edges\nunder wiki/nodes/ + per-reviewer trace under .aris/traces/comic-asset-review-loop/.The contract boundary is the asset wiki node authored by comic-asset-ref-generator (S4) at\nreview_status: \"pending\" — generated, never self-locked. This skill is pure verify + lock:
regenerate/revise it routes a ≤140-char\nrepair_instruction back to S4, which bumps the version and emits a new file.not_scored and EXCLUDED from the threshold, never substituted with 0 (the v1.0 bug; see Hard do/don't).status: \"locked\" + payload.review_status: \"locked\" on a 准×3 approve, plus the\nreview/decision/failure_mode trace. locked is the one-way door — immutable for the rest of Phase 1.Read the asset node payload (per schemas/node_schema.json, node_type: \"asset\",\npayload.required = [asset_kind, name, visual_description, identity_lock, ref_requirements, review_status, version]):
asset_kind ∈ {character, scene, prop, text_panel, logo_free_symbol} — routes the per-reviewer query\n(see §\"Asset-kind specialisation\").identity_lock.{must_preserve[], must_avoid[]} — the per-asset rubric (e.g. from ART_BIBLE.md §1:\nexecutor = blue hoodie #1D4684 + brown hair + no beard; reviewer = green hoodie #30582D + near-black\nhair + beard; \"颜色/胡子/体型轮廓不可变\").ref_requirements.{aspect_ratio, background (white|transparent), pose, isolated_reference}.output_ref.{file_path, data_url, sha256, width, height, mime} for a raster ref (all 6 fields, else schema\nviolation) OR {generator_script, owner_script, file_sha256} for a deterministic SVG/JSON source.asset-review-loop)Run before any visual reviewer call (and re-run before lock). Fail-closed; correctness costs no credits.
\ncheck_asset_collisions.py static-scans every gen_*.py with the regex\n\\bw\\(\\s*[\"']([^\"']+\\.(?:svg|json|png))[\"'] and asserts each output filename has EXACTLY ONE generator\nowner. len(writers) > 1 → exit 1. This is the tool that enforces \"one visual dialect, never two\":\ntwo generators writing the same path with different content is a silent, run-order-dependent corruption\n(whoever runs last wins). It is deliberately NOT a \"skip if file exists\" guard — that would hide the\norder bug instead of surfacing it. Verdict: exit 0 (clean) is a precondition for entering Layer 2;\nexit 1 blocks the round and routes to S4 to rename/delete the duplicate. (Verified clean at 26 outputs in\nthe worked example.)assert \"<text\" not in svg in the emitter. A zero-text asset that contains text fails the round._contract field forbids re-layout-by-eye;\nthe JSON is the single coordinate truth (e.g. wiki_starmap_nodes_v1.json for S16b labeled + S22 wordless).output_ref has all of file_path, data_url, sha256, width, height, mime; sha256 matches the file on disk (a partial write is a schema violation, not a lockable ref).Each reviewer reads the real asset file (image bytes via Read / analyzeFile) and scores all five; the\nPrimary reviewer's score is authoritative for that dim, the Secondary corroborates.
must_preserve trait, avoids EVERY must_avoid.\nPrimary = Gemini (visual), Secondary = Codex (semantic). character → face/age/hair/costume/beard;\nscene → layout + lighting-type + era.\n\nThese are THIS skill's pre-gate routing thresholds (the asset-gate step-3 routing this skill OWNS), and\nthey are written to MATCH
\ncomic-cross-layer-gate's--gate asset\nLOCK predicate verbatim. The gate'sLOCKEDcondition isidentity_lock_satisfied >= 4 AND ref_quality >= 4 AND bg_isolation >= 4 AND safety_ip >= 4AND the准 ×3same-round cross-model unanimity rule —\nreuse_readinessis ADVISORY at the gate and does not block the lock. Theapproverule below uses\nexactly those four floors;reuse_readinessonly feedsrevise/regeneraterouting, never blocks the lock.
Codex MIN-fuses each dim across the reviewers who scored it (the gate's deterministic rule: min per dim,\nmax for an inverted dim, single scorer → use it, NO scorer → not_scored EXCLUDE — never 0-substitute,\nnever average: averaging {5,3}→4 would slip a sub-floor ref past the >=4 lock; min=3 correctly fails). So\nthis inline routing can only be stricter, never weaker, than --gate asset. The legal asset verdict set is\n{approve, regenerate, locked, abandon_shot} (gate-enforced); the internal revise below is an alias that\nroutes a regenerate. Then:
approve ← identity_lock_satisfied >= 4 AND ref_quality >= 4 AND bg_isolation >= 4\nAND safety_ip >= 4 (the four gate floors). asset_kind == \"character\" override:\nidentity_lock_satisfied == 5 is REQUIRED (a raster character ref locks only at a perfect identity score —\n4 is not enough for a face hundreds of panels condition on); the other three floors stay >= 4. (准×3\nsame-round unanimity still applies on top — see the LOCK note below.)regenerate ← identity_lock_satisfied < 4 (or < 5 for a character) OR bg_isolation < 4 OR\nref_quality < 4 OR safety_ip ∈ [3, 4) OR reuse_readiness < 3 (advisory: routes a regen hint, does not\nitself block a lock that already clears the four floors).revise ← mid-band, a single targeted edit likely fixes it (softer hint, still routes a regenerate\nwith a focused repair_instruction).abandon_shot ← safety_ip < 3 (HARD STOP, IP) OR round >= MAX_REVIEW_ROUNDS without convergence.\n\nDeterministic-SVG / JSON asset branch (a parametric builder, NOT a raster ref → the visual 5-dim rubric\ndoes not apply). A deterministic-SVG source locks iff ALL FOUR hold (Layer-1 asserts + a 3-dim 0–5 read):\n
\nsingle_source_owner == pass(Layer-1check_asset_collisions.pyexit 0 — exactly one generator owns the\nfilename) ANDrender_preview_legible >= 4(the SVG/JSON rasterizes to a non-empty, on-spec preview PNG —\na RAW render pre-check, not a vote) ANDliteral_contract >= 4(the_contractfield is honored: literals\nare ascii-tokenizable, the JSON is the single coordinate truth, no eyeball re-layout) AND\nstyle_family_consistency >= 4(one visual dialect — palette/stroke/grid match theART_BIBLE.mdfamily).\nPlus the zero-text / non-gated-text asserts: a wordless source contains no<text>(assert \"<text\" not in svg); a labeled source's glyphs are exactly the contract's literals. Same 准×3 same-round unanimity gates\nthe SVG lock as the raster lock. (Gemini PRIMARY onrender_preview_legible+style_family_consistency;\nCodex PRIMARY onliteral_contract;single_source_owneris the Layer-1 tool, not a reviewer vote.)
\n\nLOCK requires 准×3 UNANIMITY, not Codex-synth-approve. Even on a Codex
\napprove, the lock fires iff\ncc_approves AND gemini_approves AND codex_approvesin the same round (each reviewer's own 5 scores\nindependently clear the per-reviewer thresholds — including theidentity_lock_satisfied == 5floor for a\ncharacter ref / the four-dim SVG floors for a deterministic source). A Codexapprovewith any reviewer below\nthreshold is DOWNGRADED torevise. (--liteis non-acquitting: with Gemini absent it may only write\nprovisionalstatus: under_review+lite_provisional: true, and can NEVER satisfy the LOCK predicate or reach P6.) This same-round\nunanimity rule is exactlycomic-cross-layer-gate's准 ×3for\n--gate asset(this skill is the OWNER of that convention) — NOT \"three consecutive rounds\".
abandon_shot (single-vote IP veto, no retries).comic-outline-creator / comic-storyboard-creator\n(the spec/outline must change; stop re-baking the image). Mutual recursion is bounded by this ceiling.face_drift, age_drift, costume_drift, pose_off, multi_subject_leak, halo_artifact, shadow_spill, non_white_bg, jpeg_artifact, glyph_break, unintended_text, watermark_present, contact_sheet, celeb_likeness, copyrighted_character, brand_logo_present, collage_grid, low_resolution, melted_anatomy. Every reject mints a\nfailure_mode wiki node fail:<asset-slug>_<tag> tagged from this set (the <tag> must match [a-z0-9_]+ —\nstrip any stray punctuation before assembling the node_id; maps to comic-director's anatomy/identity vetoes) and routes\nits tag into the S4 repair_instruction. The node MUST carry the schema-required shape (copy the fail:*\nskeleton in \"Wiki node skeletons\" below verbatim): root status: \"active\" + payload {layer, affected_shot_ids, active, repair_pattern}. Because this is an upstream asset gate with no shots yet,\naffected_shot_ids is [] (empty) — never invented. NB the shipped example failure nodes are FILES named\nfail_*.json (the colon lives only inside the node_id, e.g. fail:s09_a01) and are ALL downstream\nshot-level nodes (layer:\"panel_visual\", non-empty affected_shot_ids — fail_s09_a01.json carries\n[\"S09\"]); no asset_ref-layer example exists in the repo, so do NOT copy the example nodes for this gate —\nthe skeleton below is the authoritative shape.
cli/validate_wiki.py)Every node this skill writes obeys schemas/node_schema.json: the root carries node_id, node_type, title, status, created_at, payload (all six required), the node_id matches ^(…|review|decision|fail):[a-z0-9_-]+$,\nand the payload carries every field in validate_wiki.py PAYLOAD_REQUIRED for that type. The canonical root\nstatus for each runtime type (schema $comment; the real example nodes carry exactly these): review → \"complete\", decision → \"final\", failure_mode → \"active\". Copy these skeletons verbatim:
// review:* — one per reviewer per round (P1 cc, P2 gemini, P3 codex). node_id: review:<asset-slug>_<reviewer>_round{R}\n// PAYLOAD_REQUIRED[\"review\"] = [target_node_id, reviewer, gate_kind] (review_scores carries the 5 dims)\n{ \"node_id\": \"review:duo_canonical_ref_v001_cc_round1\", \"node_type\": \"review\", \"status\": \"complete\",\n \"title\": \"CC narrative asset review duo_canonical_ref v1 r1\", \"created_at\": \"<ISO-8601>\",\n \"payload\": { \"target_node_id\": \"asset:<asset-slug>\", \"reviewer\": \"cc\", \"gate_kind\": \"asset\",\n \"review_scores\": { \"identity_lock_satisfied\": 5, \"ref_quality\": 4, \"bg_isolation\": 4,\n \"reuse_readiness\": 4, \"safety_ip\": 5 },\n \"failure_modes_flagged\": [], \"timed_out\": false } }\n\n// decision:* — one per round after the synth+准×3 (P4). node_id: decision:asset_<asset-slug>_round{R}\n// PAYLOAD_REQUIRED[\"decision\"] = [target_node_id, verdict, gate_kind]\n{ \"node_id\": \"decision:asset_duo_canonical_ref_v001_round1\", \"node_type\": \"decision\", \"status\": \"final\",\n \"title\": \"asset_gate duo_canonical_ref v1 r1 -> locked\", \"created_at\": \"<ISO-8601>\",\n \"payload\": { \"target_node_id\": \"asset:<asset-slug>\", \"gate_kind\": \"asset\", \"verdict\": \"approve\",\n \"reasoning\": \"LOCK 准×3 — identity=5 ref=4 bg=4 safety=5 reuse=4; cc&gemini&codex approve r1\",\n \"repair_instruction\": \"\", \"prior_lock_count\": 3 } }\n\n// fail:* — only on a reject/abandon_shot (P0 Layer-1 fail, P2 parse failure-mode, P5 reject). node_id: fail:<asset-slug>_<tag>\n// PAYLOAD_REQUIRED[\"failure_mode\"] = [layer, affected_shot_ids, active] (upstream gate → affected_shot_ids: [])\n{ \"node_id\": \"fail:duo_canonical_ref_v001_costume_drift\", \"node_type\": \"failure_mode\", \"status\": \"active\",\n \"title\": \"asset_gate duo_canonical_ref costume_drift\", \"created_at\": \"<ISO-8601>\",\n \"payload\": { \"layer\": \"asset_ref\", \"affected_shot_ids\": [], \"active\": true,\n \"repair_pattern\": \"executor 蓝衣棕发无须 (must_preserve); 当前绿衣漂移 — 重生为 #1D4684 hoodie\" } }\nAfter any lock (P6), run python3 cli/validate_wiki.py examples/<project> — it MUST PASS (root status enum,\nnode_id form, payload invariants). A node missing root status or target_node_id / layer /\naffected_shot_ids / active is the exact failure the validator catches.
This is an upstream gate, so it enforces the preconditions the downstream engine fail-closes on. Do NOT\nconflate the three SVG fields (per schemas/node_schema.json + the engine): the wiki\nblueprint.payload.content_svg (top-level on the blueprint payload), the panel_spec.payload.content_blueprint\n(the panel_spec's own field — there is NO content_svg on a panel_spec), and the comic.json panel's\nruntime condition.content_svg (the field the engine reads at spiral_engine.js .condition.content_svg).
blueprint\nnode would carry it as blueprint.payload.content_svg), because at bake time the engine refuses a\ncomic.json panel whose condition.content_svg is null/absent. A ref that cannot become a content-SVG\ncondition is not \"reuse-ready\" — score reuse_readiness (and, for an SVG source, render_preview_legible)\naccordingly. Never condition.content_svg: null.condition.expected_literals. If the asset is a text_panel / figure\nsource whose downstream panels bake numbers (e.g. a wandb curve, a verdict stamp, a DDL widget), its lock\nMUST verify the literal contract holds (the literals are ascii-tokenizable and the JSON _contract forbids\neyeball re-layout) — because comic-director's gate refuses to run a baked figure-panel whose comic.json\ncondition.expected_literals is empty. Locking a figure source that cannot supply gateable literals would\nstrand the panel; a scene panel with no audited numbers takes text_mode: \"html\" instead.P0 — Resolve + Layer-1 gate. Resolve the target asset node (single asset_id, or fan out one branch per\nasset under --batch-from-outline). Assert review_status == \"pending\"; abort if already locked (one-way\ndoor). Route by asset_kind (visual raster kinds → the raster 5-dim branch; a deterministic SVG/JSON source →\nthe SVG 4-dim branch; an audio kind would take the audio branch — out of scope for the comic). Run Layer 1\n(check_asset_collisions.py + build-asserts + raster 6-field check); on any failure, write a failure_mode\nnode (root status: \"active\", payload {layer:\"asset_ref\", affected_shot_ids:[], active:true, repair_pattern: <the Layer-1 tag-derived hint>} per the fail:* skeleton) + a failure_of edge fail:<…> → asset:<asset-slug>,\nroute to S4, and stop — do not spend a visual reviewer call.
P1 — Claude narrative review. Read the PNG (image bytes). Score all 5 dims (Claude PRIMARY on\nreuse_readiness; Codex is PRIMARY on safety_ip per the rubric), emit failure_modes_flagged[] (from the\nfixed vocab) + narrative_notes. Write a review-stage scratch JSON and a wiki review node\nreview:<asset-slug>_cc_round{R} with root status: \"complete\" and payload carrying the three\nPAYLOAD_REQUIRED fields target_node_id: \"asset:<asset-slug>\", reviewer: \"cc\", gate_kind: \"asset\" plus\nreview_scores keyed by the 5 dims (see the review:* skeleton above). Append a reviews edge\nreview:<…> → asset:<asset-slug> (review → target — the direction the gate walks; NEVER reviewed_by).\nAtomic write (.tmp → mv).
P2 — Gemini visual review (skip iff --lite). Call mcp__gemini__analyzeFile with model: \"auto-gemini-3\" (NEVER raw gemini-3.1-pro), the PNG path, and a strict-JSON-only template that asks for\nthe 5 dim scores (Gemini PRIMARY on identity_lock_satisfied, ref_quality, bg_isolation) +\nfailure_modes drawn from the fixed vocab. Retry once on non-JSON, then log a gemini_parse_failure and\ntreat those scores as null (not 0). Optional focused follow-up if identity_lock_satisfied < 3. Write the\nreview node review:<asset-slug>_gemini_round{R} with root status: \"complete\" and payload\ntarget_node_id: \"asset:<asset-slug>\", reviewer: \"gemini\", gate_kind: \"asset\" + review_scores; append\nthe reviews edge review:<…> → asset:<asset-slug> (review → target). On a parse failure mint a fail:*\nnode per the fail:* skeleton (root status: \"active\", payload {layer:\"asset_ref\", affected_shot_ids:[], active:true, repair_pattern:\"gemini_parse_failure — re-elicit strict JSON\"}).
P3 — Codex synth gate. mcp__codex__codex (r1) / mcp__codex__codex-reply (r2+ with the saved\nthreadId), effort xhigh, no model pin (follows the local codex config — currently gpt-5.6-sol). Pass\nONLY numeric scores + failure_modes + asset_metadata + file paths
narrative_notes/prose (reviewer-independence;\na prose leak ABORTS the call with a loud error). Codex adds its OWN safety_ip (PRIMARY) + semantic\nidentity_lock_satisfied + reuse_readiness scores. First persist Codex's OWN reviewer node (mirroring P1\ncc / P2 gemini): review:<asset-slug>_codex_round{R} (root status: \"complete\", payload {target_node_id: \"asset:<asset-slug>\", reviewer: \"codex\", gate_kind: \"asset\", review_scores: {the 5 dims Codex scored}}) + a\nreviews edge review:<…>_codex_round{R} → asset:<asset-slug> — so all THREE reviewer families exist as\nfirst-class collected nodes (准×3 needs Codex as one of the three the gate walks). THEN synthesise: MIN-fuse\neach dim across the reviewers who scored it (max for an inverted dim; single scorer → use it; no scorer →\nnot_scored, EXCLUDE — never average / 0-substitute) → verdict + per_reviewer_unanimity{cc,gemini,codex} +\nrepair_instruction (≤140 chars) + rationale (≤300 chars — the ONLY Codex prose allowed out).P4 — Apply threshold + 准×3 → decision node. Apply the threshold, then the unanimity downgrade. Write a\ndecision node decision:asset_<asset-slug>_round{R} with root status: \"final\" and payload carrying the\nthree PAYLOAD_REQUIRED fields target_node_id: \"asset:<asset-slug>\", verdict, gate_kind: \"asset\" plus\nreasoning, repair_instruction, prior_lock_count (see the decision:* skeleton above). approve +\nprior_lock_count >= 3 (准×3) → mark lock. approve without unanimity → downgrade to revise, which is\npersisted as verdict: regenerate — the only legal decision.payload.verdict tokens are\n{approve, regenerate, locked, abandon_shot} (a verdict outside the gate's set is a hard error); revise is an\ninternal route label kept in reasoning/repair_instruction, never written as the verdict. A TERMINAL\nfail — abandon_shot (IP veto) OR the R == MAX_REVIEW_ROUNDS escalation (P5) — sets BOTH root\nasset.status: \"rejected\" AND payload.review_status: \"rejected\" (canon D2/D3: terminal-fail → root status\nrejected; mirror the lock path, which sets both). A non-terminal revise/regenerate keeps root\nstatus: \"under_review\" (provisional), never leaving an abandoned asset stuck at pending.
\n\nGate handoff (this skill is the asset gate's review-producer). P1–P3 have now persisted the
\nreview:*\nscore-nodes +reviewsedges thatcomic-cross-layer-gate--gate assetcollects — so the gate is never invoked cold. This skill runs the asset-gate synthesis inline (it\nOWNS 准×3); when the orchestrator defers acquittal tocomic-cross-layer-gate <asset:<asset-slug>> --gate assetinstead, it fuses the SAMEreview:*nodes via thosereviewsedges and re-derives the>=4-floor +\n准×3 verdict. NOTE this skill applies an ADDITIONAL stricter pre-gate floor (character\nidentity_lock_satisfied == 5; the deterministic-SVG 4-dim branch for an identity-4 ref), so its inline\nverdict can be stricter (never weaker) than the bare gate's>=4floor — the deferred gate does not\nreproduce the==5character rejection byte-for-byte. Thedecidesedge (P5) links thisdecision→ the asset.
P5 — Edges + route. Write edges using ONLY validate_wiki.py EDGE_TYPES: a reviews edge\nreview:<…> → asset:<asset-slug> per reviewer (review → target — the direction --gate asset walks; NEVER\nreviewed_by, the opposite direction, which the gate cannot follow), the decides edge\ndecision:asset_<…> → asset:<asset-slug>, and on a reject/abandon_shot a failure_of edge\n{src: \"fail:<asset-slug>_<tag>\", dst: \"asset:<asset-slug>\", type: \"failure_of\"} (failure_mode → target,\nmatching the real examples/comic_m3_audit/wiki/edges.jsonl). A regen that supersedes the prior version writes a\nsupersedes edge asset:<new> → asset:<old> (bare node_ids, no @vN). Then route:
--lite) and NO\nfamily hard-failed a floor → RE-VOTE: re-collect ONLY the missing reviewer (re-run that one family); do\nNOT --regenerate (re-baking wastes a credit when the image is fine and only the third vote is missing).regenerate/revise (a family HARD-FAILED a floor) and R < 3 → call\ncomic-asset-ref-generator --regenerate with the\nrepair_instruction (it bumps the version, emits a new PNG with a supersedes self-edge, re-fires at R+1).regenerate/revise and the re-bake count reaches MAX_ASSET_REGEN = 4 (the gate's re-bake cap,\ncomic-cross-layer-gate lines 70/184 — a SEPARATE ceiling from this\nskill's own MAX_REVIEW_ROUNDS = 3 visual review-round limit; either hitting its ceiling escalates) → force\nabandon_shot → ESCALATE the requirement to outline/storyboard.abandon_shot (IP) → exit immediately.P6 — Lock (on approve + full 准×3 = CC AND Gemini AND Codex same-round unanimity). Set status: \"locked\"
payload.review_status: \"locked\"; append tags [asset_gate_locked, round_{R}_approve]. (--lite can NOT\nreach P6 — with Gemini down it stays provisional under_review, never locked.) Validate against\nschemas/node_schema.json (python3 cli/validate_wiki.py examples/<project>); append a log.md line; write\nthe final ASSET_REVIEW.md score_progression table (the human audit trail — why it locked at r2 vs r3\nvs abandoned). Lock = ONE-WAY DOOR, immutable for the rest of the pipeline.ART_BIBLE.md §1 iron law:\nexecutor never bearded, reviewer always bearded; colours immutable, L/R position free).bg_isolation becomes scene_emptiness (a location plate must carry no characters/props).bg_isolation weighted (clean cutout, no spill).ref_quality (maps to comic-director's\nbaked_text_quality); the literal contract (engine contract #2) is checked at lock.safety_ip weight (the symbol must not resemble a real brand mark).IDENTICAL scores across review rounds mean the rubric/judge is broken — STOP regenerating and audit the\njudge. If round R and round R+1 (on a regenerated asset) return the same per-dim vector from the same\nreviewer, the artifact changed but the score did not → the rubric (not the asset) is wrong. Concretely:
\ncc/gemini/codex\nreturn a byte-identical 5-vector on a new file (different sha256), flag judge_stuck.--regenerate. Two identical rounds is the ceiling for grinding\neven before MAX_REVIEW_ROUNDS.identity_lock.{must_preserve,must_avoid} and the ART_BIBLE.md carve-\nouts — is the gate flagging a by-design trait as drift? (The classic: the two-world warm/dark palette is\nintentional — ART_BIBLE.md §0.5 \"冷暖是 by-design 的,不是漂移\"; a naive gate scores it \"style drift\"\nevery round. The motif-vs-env disambiguation is the same lesson: the gate must know which instances are\ncontinuity-bearing.) Fix the rubric (sharpen the per-dim definition / add the carve-out to the reviewer\nquery), record it as a decision trace node, then resume. A stuck score is a judge bug, not an asset bug.Ground every run in the real M3-audit comic asset set under examples/comic_m3_audit/ (repo-relative). The exact patterns to copy:
Layer-1 collision gate — gen/check_asset_collisions.py (41 lines) is the single-source CI gate\nverbatim: the WRITE = re.compile(r'\\bw\\(\\s*[\"']([^\"']+\\.(?:svg|json|png))[\"']') scan over every gen_*.py,\nthe len(writers) > 1 collision rule, the exit 0/1 contract, and the load-bearing comment that it is\ndeliberately NOT a \"skip if the file already exists\" guard (\"that would HIDE the order bug instead of\nsurfacing it\"). Its message — \"Each asset must have ONE canonical owner ('one visual dialect, never two')\"\n— is the literal embodiment of this skill's single-source rule. Verified clean at 26 outputs (exit 0).\nRun this as Layer 1 of P0; do not enter Layer 2 until it returns 0.
The identity-lock rubric — ART_BIBLE.md §1 (角色身份锁) is the literal must_preserve/must_avoid\nsource for an identity_lock_satisfied score: authority ref duo_canonical_ref_v001.png; per-character hex\ntable (executor blue hoodie #1D4684 / brown hair #925935 / NO beard; reviewer green hoodie #30582D\n/ near-black hair #0F0D16 / beard); the 铁律 \"executor 永远蓝+棕发+无须;reviewer 永远绿+黑发+有须 …\n颜色/胡子/体型轮廓不可变\". A Gemini visual reviewer cat's the PNG and scores it against this — not by\nvibes. §0.5 (the warm/dark two-world palette, \"冷暖是 by-design 的,不是漂移\") is the canonical carve-out\nfor the judge-is-broken guard above. §7 FORBIDDEN (no smooth gradients on character surfaces, no voxel, no\nnon-canonical mascot colours, no off-identity beard) is the must_avoid list.
The P0-blocker pattern — story/STORYBOARD_DRAFT.md flags researcher_chibi_canonical_ref_v001.png as\nmissing (only the duo had a canonical ref). That is exactly the state this gate must catch before any\npanel bakes the researcher: a raster ref with no file is not lockable (Layer-1 raster 6-field check fails);\nit routes to S4 to generate the ref, and only then does Layer 2 run. Copy this: a panel that conditions\non a ref whose asset node is not locked is a pre-bake blocker, surfaced here, never discovered downstream.
Pattern to copy end-to-end: Layer 1 (deterministic, free) gates single-source + completeness → Layer 2\n(cross-model, metered only after L1 passes) gates the 5 dims against the ART_BIBLE.md rubric → lock at 准×3\nor route back to S4 (regenerate) / outline (escalate).
check_asset_collisions.py) as Layer 1 of every run, before any\nreviewer call — \"one visual dialect, never two\" is enforced by the tool, not by prose.approve alone\ndowngrades to revise. The generator family never self-acquits its own ref.safety_ip < 3 from any one reviewer as an immediate hard abandon_shot — IP is a single-vote\nveto, no retries (the same shape as the anatomy single-vote veto).not_scored and EXCLUDED from the threshold —\nsubstituting 0 was the v1.0 bug that mis-rejected good refs.regenerate/revise, route a ≤140-char repair_instruction to\ncomic-asset-ref-generator; this skill only verifies and locks.locked asset — the lock is a one-way door; a later need for a different ref is a NEW\nasset (new file, supersedes edge), not an edit of the locked one.| Target | \nFit | \n
|---|---|
| reusable character sheet / location plate / prop cutout (identity-locked raster ref) | \nexcellent | \n
| deterministic SVG/JSON asset source (one parametric builder per recurring motif) | \nexcellent | \n
| text_panel / figure source whose downstream panels bake gateable literals | \ngood (checks the literal contract) | \n
| logo-free symbol / mark (extra safety_ip weight) | \ngood | \n
| a composed panel's faithfulness | \nno → that's comic-director's panel_gate | \n
| cross-panel continuity (motif drift across the film) | \nno → that's comic-continuity-audit / the assembly_gate | \n
reviewer-independence — each reviewer reads the real asset file; the Codex synth gets numeric scores + failure_modes + paths only, never another reviewer's prose (a leak ABORTS).acceptance-gate — the loop DRIVES (regenerate, re-fire R+1) but can't ACQUIT: the LOCK is a Type-B verdict that needs 准×3 (Gemini + Codex + Claude in the same round); Codex-synth-approve alone is downgraded; same-family N-of-Claude is never a jury.review-tracing — every Codex/Gemini reviewer call's full prompt+response is saved under .aris/traces/comic-asset-review-loop/<date>_run<NN>/, so each lock/abandon_shot verdict is auditable; the ASSET_REVIEW.md score_progression is the human-readable trail.reviewer-routing — Codex reviewer calls pin no model (local codex config — currently gpt-5.6-sol) at effort xhigh (only the upstream bake payload pins gpt-5.5 + xhigh, the run_comic.get_bake_plan() compat default; config-driven override planned); Gemini auto-gemini-3 (never raw gemini-3.1-pro); — reviewer: oracle-pro routes the synth to mcp__oracle__consult (the Pro tier) on explicit request; never downgrade the tier.