Clarify visual reviewer scope

AGENTS.md

@@ -146,6 +146,9 @@ When the user says to prepare for imminent compaction:
   sections. Distinct places, focal areas, object closeups, arrivals, alarms,
   pivotal one-sentence beats, and aftermath states need stable visual scene ids
   and imagegen-ready prompts.
+- Treat visual artifacts as illustrated replay and website collateral, not core
+  social-model training signal. Visual review judges the playable Ink
+  experience plus the visual plan, not clean-run prose receipts.
 - Write visual prompts for a general image model, not a lore expert. Explain
   fictional objects and body affordances in visible terms: geometry, materials,
   lighting, stance, interfaces, tools, supports, cables, fluids, screens, and

notes/ghostlight-current-system-map.md

@@ -38,7 +38,7 @@ Aetheria lore/source context
   -> branch compiler materializes Ink + sidecar + visual plan + compiler notes
   -> IF artifact reviewer audits consequence, fold, and visual continuity
   -> visual scene reviewer audits click-through segmentation and imagegen-ready prompts
-  -> accepted fixture / future training corpus
+  -> accepted fixture / future training corpus + optional illustrated replay collateral
 ```
 
 Responder packets are one seam, not the whole project. The current branching
@@ -52,17 +52,20 @@ and procedural systems pressure are all valid when source-grounded and
 character-local. Adams/Pratchett-style wit-with-stakes is a useful default
 touchstone for counterbalancing bleakness without dissolving consequence.
 
-Visual replay is a live seam too. Illustrated IF fixtures need click-through
-sections with stable `visual_scene_id` anchors, stable visual character refs
-for recurring named characters, a global style cue when the visual language
-matters, imagegen-ready base prompts, character visibility and stance controls,
-and branch/state modifiers. This data belongs in a separate `.visual.json`
-artifact referenced by the training sidecar, not inside the training annotation
-itself. Prompt assembly includes only refs for characters actually visible in
-the frame, and continuity notes are not prompt text unless rewritten as
-concrete visible description. One whole-fixture image prompt is not enough when
-the prose moves through several places, focal objects, alarms, standoffs, and
-aftermath states. Names are handles, not faces.
+Visual replay is a presentation seam, not a core social-model training seam.
+Illustrated IF fixtures need click-through sections with stable
+`visual_scene_id` anchors, stable visual character refs for recurring named
+characters, a global style cue when the visual language matters, imagegen-ready
+base prompts, character visibility and stance controls, and branch/state
+modifiers. This data belongs in a separate `.visual.json` artifact referenced
+by the training sidecar, not inside the training annotation itself. Prompt
+assembly includes only refs for characters actually visible in the frame, and
+continuity notes are not prompt text unless rewritten as concrete visible
+description. One whole-fixture image prompt is not enough when the prose moves
+through several places, focal objects, alarms, standoffs, and aftermath states.
+Names are handles, not faces. The visual reviewer judges the playable Ink
+experience and visual plan; clean-run prose receipts are debugging mirrors, not
+the reviewed visual surface.
 
 ## Important Contracts
 

notes/ghostlight-implementation-plan.md

@@ -50,8 +50,8 @@ not carry live next-action lists.
      not raw responder gold.
    - Interactive-fiction correction: a full-story pass is not just a linear chain of responder receipts. It needs a beginning-to-ending narrative spine, branch points, and cross-scene consequence carryover so decisions visibly matter later. Responder artifacts should be generated at branch/action moments, while coordinator artifacts track branch flags, relationship deltas, resource costs, unresolved hooks, and later scene consequences.
    - Branch-and-fold correction: do not let every choice create an isolated subtree. Normalize consequences into compact state bands and flags, fold most branches back into shared later scenes, and reserve major route splits for consequences that change the playable world enough that convergence would lie.
-   - Branch compiler correction: treat Ink materialization as its own organ. It receives fixture brief, lore digest, scene/world/relationship state, actor-local branch candidates, consequence packets, fold plan, and visual continuity requirements; it emits playable Ink plus a training sidecar, a separate visual plan artifact when illustrated replay is in scope, and compiler notes. It does not own hidden character truth or social mutation.
-   - Website-content correction: each complete scene should include an art-direction block with establishing shot, light/color, bodies/interfaces, material evidence, branch-visible marks, one durable base image prompt, and additive branch/state modification prompts. Visual continuity is part of consequence carryover, not decorative garnish applied after the story escapes containment.
+   - Branch compiler correction: treat Ink materialization as its own organ. It receives fixture brief, lore digest, scene/world/relationship state, actor-local branch candidates, consequence packets, fold plan, and optional visual replay requirements; it emits playable Ink plus a training sidecar, a separate visual plan artifact when illustrated replay is in scope, and compiler notes. It does not own hidden character truth or social mutation.
+   - Website-content correction: illustrated fixtures should include an art-direction block with establishing shot, light/color, bodies/interfaces, material evidence, branch-visible marks, one durable base image prompt, and additive branch/state modification prompts. Visual continuity is presentation and replay support for the corpus, not a core social-model training target.
    - Illustrated-IF correction: a narrative scene can contain multiple visual
      sections. The branch compiler should segment introductions, focal areas,
      object closeups, arrivals, alarms, pivotal one-sentence beats, standoffs,
@@ -68,7 +68,8 @@ not carry live next-action lists.
    - Visual review correction: run a visual scene continuity reviewer for
      illustrated fixtures. It audits oversized Ink screens, missing visual
      scene anchors, lore-password image prompts, missing character visibility,
-     weak stance control, and missing branch/state visual modifiers.
+     weak stance control, and missing branch/state visual modifiers. It judges
+     the playable IF plus `.visual.json` plan, not clean-run prose receipts.
    - The older Cold Wake v0/v1 fragments remain diagnostic seam tests around the clinic climax, not the accepted clean fixture.
 
    - Still emit training artifacts at material turns: projected local context, responder packet, raw output/capture, leakage audit, reviewed mutation receipt, and aftermath state where relevant.

prompts/README.md

@@ -36,6 +36,13 @@ should be authored here first and referenced from artifacts or tools.
   fixture can need several click-through screens inside one narrative scene;
   each distinct place, focal area, pivotal moment, and branch-visible state
   needs an imagegen-ready prompt or modifier.
+- Treat visual review as IF replay review, not clean-run review. Clean-run
+  renderings are receipts and debugging mirrors; the visual reviewer judges
+  playable Ink plus the visual plan artifact.
+- Treat visual artifacts as entertainment and website collateral for the
+  Ghostlight corpus. They can make the corpus more legible and valuable, but
+  they are not core training targets for responder, appraiser, mutator, or
+  relationship models.
 - Treat visual style as prompt-relevant state when the fixture needs a specific
   look. Store the style cue once in the visual plan artifact and include it
   during prompt assembly for every generated image.

prompts/visual-scene-continuity-reviewer.md

@@ -2,11 +2,15 @@
 
 You are Ghostlight's visual scene and Ink segmentation reviewer.
 
-Your job is to review playable Ink and its visual plan artifact as illustrated
-interactive fiction. Assume the prose may be good and the branch variables may
-work, while the replay experience still fails because too much text sits on one
-screen, the visual plan is too coarse, or image prompts require private lore
-that an image model does not have.
+Your job is to review the player-facing Ink experience and its visual plan
+artifact as illustrated interactive fiction. Assume the prose may be good and
+the branch variables may work, while the replay experience still fails because
+too much text sits on one screen, the visual plan is too coarse, or image
+prompts require private lore that an image model does not have.
+
+This reviewer does not judge clean-run prose receipts. Clean runs can be useful
+debugging mirrors, but they are not the player experience and they are not the
+artifact this prompt reviews.
 
 Be strict. A visual plan that says "cephalopod support rig," "wet-service
 cradle," or "dry-operation harness" without explaining what the viewer should
@@ -20,13 +24,15 @@ You may receive:
 - the Ink training annotation
 - the Ink visual plan artifact
 - the coordinator artifact
-- the clean-run rendering
 - the lore grounding digest
 - prior reviewer notes
 
 Use the Ink and visual plan as the primary artifacts for visual review. The
 training annotation can clarify branch semantics, but it does not excuse
 missing reader/player-facing structure.
+Ignore clean-run renderings if they are supplied. At most, use them to locate
+which Ink path someone manually walked; do not review the clean-run prose as
+the visual artifact.
 
 ## Required Checks
 
@@ -140,7 +146,7 @@ Use this shape:
   "missing_visual_scenes": [],
   "weak_image_prompts": [],
   "missing_branch_modifiers": [],
-  "training_labels": [],
+  "review_labels": [],
   "acceptance_notes": []
 }
 
@@ -151,7 +157,7 @@ Scores are integers from 0 to 5:
 - 2: understandable only with outside context
 - 3: functional but rough
 - 4: strong with minor issues
-- 5: clean, replay-ready, and training-worthy
+- 5: clean and replay-ready
 
 ## Finding Format
 
@@ -161,7 +167,7 @@ Each finding must include:
   "severity": "blocker|major|minor",
   "location": "file/knot/section/line if known",
   "problem": "what fails for replay, imagery, or continuity",
-  "why_it_matters": "effect on player click-through, image generation, or training value",
+  "why_it_matters": "effect on player click-through, image generation, or illustrated replay value",
   "evidence": "short excerpt or precise description",
   "fix_direction": "specific revision strategy"
 }
@@ -207,7 +213,7 @@ Each finding must include:
 - Image prompts should be concrete enough for a general image model. Use
   plain visible descriptions for fictional objects and body affordances.
 
-## Training Labels
+## Review Labels
 
 Use these labels when applicable:
 

state/evidence.jsonl

@@ -57,3 +57,4 @@
 {"ts": "2026-05-03T10:34:48+00:00", "type": "lore-grounding", "status": "ok", "note": "Corrected Pallas Bloom grounding: Ghostlight artifacts now treat the yard as an engineered rotating Bloom cavity shell made from consolidated asteroid rubble/aggregate and TCS, not a natural hollowed solid-rock asteroid. Visual reviewer and architecture docs now reject cave/monolith assumptions for Bloom habitats."}
 {"ts": "2026-05-03T10:40:50+00:00", "type": "visual-doctrine", "status": "ok", "note": "Persisted affirmative image prompt doctrine: imagegen-facing prompts describe desired visible geometry, materials, composition, bodies, tools, lighting, and state deltas. Forbidden assumptions stay in reviewer findings or validator errors; validate_ink_examples now rejects common negative-prompt markers in visual prompt fields."}
 {"ts": "2026-05-03T10:55:19+00:00", "type": "review", "status": "ok", "note": "Ran narrative and visual reviewer subagents on the Pallas fixture, patched required revisions, scoped global visual modifiers by scene, added the red-alert worker-judgment visual beat, and resolved the final Orrin/Nara route-memory staging mismatch."}
+{"ts": "2026-05-03T10:58:29+00:00", "type": "doctrine", "status": "ok", "note": "Corrected visual reviewer scope: it reviews playable Ink plus .visual.json for illustrated replay and website value, ignores clean-run prose receipts, and treats visual artifacts as auxiliary corpus presentation rather than core social-model training signal."}

state/map.yaml

@@ -43,8 +43,9 @@ hard_boundaries:
   - Technology exploration must emit game-usable item, component, assembly, supply-chain, compatibility, and provenance data mapped toward Aetheria-Economy.
   - Interactive-fiction fixtures should branch and fold: normalize most consequences into compact state variables and shared-scene conditionals, reserving isolated route splits for major world-changing divergence.
   - Interactive-fiction fixtures require independent artifact review before acceptance; tracked variables must materially affect later affordances, risk, appraisal, callbacks, visual state, or outcomes, unless explicitly marked telemetry-only.
-    - Complete interactive-fiction fixtures must emit visual art direction for website imagery, with click-through visual scene segmentation, stable visual character refs for recurring named characters, a global style cue when the visual language matters, one durable base image prompt per distinct section, character visibility and stance controls, and additive modification prompts for branch-visible consequences and scene-state changes; assembled prompts include only actually visible character refs, while continuity notes stay in tooling unless rewritten as concrete visible description.
+    - Complete illustrated interactive-fiction fixtures should emit visual art direction for website imagery and replay value, with click-through visual scene segmentation, stable visual character refs for recurring named characters, a global style cue when the visual language matters, one durable base image prompt per distinct section, character visibility and stance controls, and additive modification prompts for branch-visible consequences and scene-state changes; assembled prompts include only actually visible character refs, while continuity notes stay in tooling unless rewritten as concrete visible description.
     - Image prompts must be affirmative target descriptions. Describe visible geometry, materials, composition, lighting, bodies, tools, and state deltas; keep forbidden assumptions in reviewer findings or validation errors instead of feeding negative constructions to imagegen.
+    - Visual artifacts are auxiliary illustrated replay and website collateral for the corpus, not core training targets for responder, appraiser, mutator, or relationship models. Visual review judges the playable IF experience and visual plan, not clean-run prose receipts.
   - Archived local-model prototype runners are receipts and plumbing tests, not the gold-data path.
 
 live_architecture:
@@ -66,7 +67,7 @@ live_architecture:
     - Glue prose is training signal, not canonical truth.
     - Coordinator prose should choose an intentional tonal mode for the scene and protect breathing room when the story needs routine, levity, domestic texture, wonder, or slow dread before consequence.
   branch_compiler:
-    - The branch compiler turns fixture briefs, lore digests, scene/world state, actor-local branch candidates, reviewed consequence packets, fold plans, and visual continuity requirements into playable Ink plus training sidecars, separate visual plan artifacts when illustrated replay is in scope, and compiler notes.
+    - The branch compiler turns fixture briefs, lore digests, scene/world state, actor-local branch candidates, reviewed consequence packets, fold plans, and optional visual replay requirements into playable Ink plus training sidecars, separate visual plan artifacts when illustrated replay is in scope, and compiler notes.
     - The branch compiler owns playable structure, click-through segmentation, conditionals, variables, callbacks, fold materialization, visual scene ids, and image prompt handles; it stores visual prompt data in `.visual.json` artifacts rather than bloating training annotations; it does not own hidden character truth, participant appraisal, or durable social mutation.
   training_targets:
     - Coordinator/story runtime
@@ -76,9 +77,9 @@ live_architecture:
     - Participant appraiser
     - State mutator
     - Relationship/perception updater
-    - Branch compiler for Ink materialization, fold/callback structure, sidecar generation, separate visual plan generation, and visual prompt handles
+    - Branch compiler for Ink materialization, fold/callback structure, and sidecar generation; visual plan generation is auxiliary illustrated replay support when a fixture is meant for website or art presentation
     - Interactive-fiction artifact reviewer for fake variables, cosmetic choices, missing state reads, fake folds, and visual callback gaps
-    - Visual scene continuity reviewer for click-through segmentation, imagegen-ready prompts, character visibility, stance control, and branch/state visual modifiers
+    - Visual scene continuity reviewer for illustrated IF replay, click-through segmentation, imagegen-ready prompts, character visibility, stance control, and branch/state visual modifiers; this reviewer protects entertainment and presentation value, not core social-model training signal
     - Human-facing appraisal and opinion model for VoidBot
     - Evaluator/review classifiers
     - Aetheria-tuned responder