Add Cold Wake story lab experiment

docs/architecture/projection-distillation-plan.md

@@ -10,7 +10,7 @@ This is a staged plan, not an immediate fine-tune order.
 ## Goal
 
 Train or adapt a smaller projection model that can turn structured Ghostlight
-state into compact speaker-local pressure artifacts.
+state into compact character-local pressure artifacts.
 
 Input:
 
@@ -26,12 +26,14 @@ Output:
 - active pressures
 - tensions
 - response affordances
+- physical or action affordances when relevant
 - voice surface
 - rendered prompt text
 - audit notes or trace linking outputs back to state
 
-The student projector should not generate final dialogue. Its job is to build
-the pressure model that a dialogue model can write from.
+The student projector should not generate final prose. Its job is to build the
+pressure model that a response model can write from. The response may be
+dialogue, action, silence, withdrawal, violence, or a mixed beat.
 
 ## Aetheria Training Boundary
 
@@ -107,7 +109,7 @@ mirror for inspection, review, and discussion.
     "scene_id": "scene-broken-taxi-oz",
     "speaker_agent_id": "cat_marrigan",
     "listener_ids": ["oz_operator"],
-    "projection_mode": "dialogue_turn"
+    "projection_mode": "response_turn"
   },
   "output": {
     "known_facts": [],
@@ -179,8 +181,10 @@ For each projection, check:
   scene pressure?
 - Did response affordances emerge from state intersections?
 - Did voice surface stay separate from defensive behavior?
+- Did the projection allow nonverbal or physical action when the state and scene
+  made speech less likely than movement?
 - Did known secrets become guarded pressure rather than forbidden context?
-- Did the rendered prompt give the dialogue model enough slope to produce the
+- Did the rendered prompt give the response model enough slope to produce the
   target behavior without commanding the exact behavior label?
 
 ## Fine-Tune Readiness
@@ -214,5 +218,6 @@ Minimum useful milestone:
 ## Boundaries
 
 The student projector is not the character. It is not the dialogue writer. It is
-not the evaluator. It is a compiler from structured social state to compact
-speaker-local pressure. Keep those responsibilities separate.
+not the action choreographer. It is not the evaluator. It is a compiler from
+structured social state to compact character-local pressure. Keep those
+responsibilities separate.

docs/architecture/prompt-projection-contract.md

@@ -1,34 +1,38 @@
 # Prompt Projection Contract
 
-Ghostlight should not prompt a dialogue model with the full schema. The schema is
+Ghostlight should not prompt a response model with the full schema. The schema is
 state storage. A prompt is a temporary scene instrument.
 
-The projection layer turns durable state into a compact speaker-local pressure
+The projection layer turns durable state into a compact character-local pressure
 model, then into prompt prose. The prose should describe what the character
-knows, wants, fears, protects, misreads, and sounds like right now. It should not
-recite taxonomy labels unless a debugging mode explicitly asks for them.
+knows, wants, fears, protects, misreads, can physically do, and sounds like right
+now. It should not recite taxonomy labels unless a debugging mode explicitly asks
+for them.
 
 ## Projection Pipeline
 
 ### 1. Select Speaker-Local Inputs
 
-For one speaker in one scene, gather only the state that can affect this turn:
+For one acting character in one scene, gather only the state that can affect this
+turn:
 
-- speaker identity and public role
-- speaker canonical state
-- speaker goals and protected values
-- speaker memories selected by scene, relationship, and active pressure
-- speaker perceived overlay of the listener
-- speaker-to-listener relationship stance
-- listener-to-speaker relationship stance only if the speaker has reason to infer it
+- character identity and public role
+- character canonical state
+- character goals and protected values
+- character memories selected by scene, relationship, and active pressure
+- character perceived overlay of the listener or target
+- character-to-listener relationship stance
+- listener-to-character relationship stance only if the character has reason to infer it
 - scene public facts
-- scene hidden facts known to the speaker
+- scene hidden facts known to the character
 - active public and private stakes
 - culture, faction, or institution priors relevant to the scene
+- physical affordances, constraints, distance, tools, weapons, bodies, access,
+  and social permissions relevant to possible action
 
 Unknown or author-only material is omitted. Known secrets are included only as
-speaker-local pressure: what the speaker is guarding, why it matters, and what
-would happen if it leaked.
+character-local pressure: what the character is guarding, why it matters, and
+what would happen if it leaked.
 
 ### 2. Compute Active Pressure
 
@@ -73,8 +77,23 @@ Tensions are where verbal behavior starts to move.
 
 ### 4. Derive Response Affordances
 
-Do not store defensive speech moves as canonical traits. Derive likely response
-affordances from state intersections.
+Do not store defensive speech moves or action moves as canonical traits. Derive
+likely response affordances from state intersections.
+
+Responses can be:
+
+- speech
+- action
+- silence
+- withdrawal
+- violence
+- mixed speech and action
+
+The renderer must not assume a line of dialogue is the right output. Sometimes
+the right next move is leaving the room, taking someone's wrist off a console,
+opening the airlock interlock, refusing to answer, or slapping someone in the
+face because the relationship, status threat, bodily proximity, and inhibition
+profile finally made language too slow.
 
 Examples:
 
@@ -85,6 +104,9 @@ Examples:
 | Emotional evasion | Distance seeking + withdrawal + acute shame or control pressure + low self-disclosure. |
 | Banter as boundary | Humor + ironic distance + attachment pressure + distance seeking. |
 | Verbal aggression | Hostility + grievance activation + perceived status threat + abrasive boundary. |
+| Physical withdrawal | Threat sensitivity + distance seeking + low trust + available exit + low obligation to stay. |
+| Coercive interruption | Control pressure + urgency + authority role + high stakes + reachable control surface. |
+| Slap or shove | Acute shame or moral disgust + proximity + low inhibition + perceived violation + limited verbal trust. |
 
 These are test targets, not prompt commands. The rendered prompt should describe
 the pressure in ordinary language:
@@ -96,9 +118,18 @@ part of the question while making the emotional part sound ridiculous.
 
 That gives the model a behavioral slope without saying `use sarcastic_deflection`.
 
+For action, use the same rule:
+
+```text
+If Veyr calmly describes a possible person as a diagnostic residue while Sella
+is close enough to touch him, Sella may stop negotiating and physically break
+the rhythm of the room. This should be rare, costly, and grounded in her care
+ethic under overload, not random spice.
+```
+
 ### 5. Render Prompt Sections
 
-A dialogue prompt should be compact and scene-bound.
+A response prompt should be compact and scene-bound.
 
 Recommended sections:
 
@@ -112,8 +143,10 @@ Recommended sections:
 - `Active Inner Pressures`: the strongest motives, fears, wounds, and needs
 - `Relationship Read`: how the speaker currently interprets the listener
 - `Tensions`: contradictions the speaker is managing
+- `Action Affordances`: what the speaker can physically do and what it would
+  cost
 - `Voice Surface`: baseline speech texture for this moment
-- `Likely Defensive Moves`: ordinary-language pressure outcomes, not schema labels
+- `Likely Response Moves`: ordinary-language pressure outcomes, not schema labels
 - `Do Not Invent`: boundaries about facts absent from the pack
 - `Task`: the narrow generation request
 
@@ -128,15 +161,20 @@ will not receive.
 
 Generation and evaluation are separate passes.
 
-The dialogue model receives the rendered prompt and writes from the speaker's
+The response model receives the rendered prompt and writes from the character's
 local pressure model. A later evaluator can check whether target emergent
-behaviors appeared when the state predicted them.
+behaviors appeared when the state predicted them. The output may be dialogue,
+action, silence, or a beat combining them.
 
 Evaluation questions can include:
 
 - Did the line protect the speaker's known secret without naming unknown facts?
 - Did the line reflect the active relationship stance?
 - Did the line preserve the voice surface without flattening into a generic bit?
+- If the output used action, was the action justified by physical affordances,
+  relationship pressure, inhibition, and stakes?
+- If the output avoided speech, was silence or movement more plausible than
+  another line?
 - Did a target defensive behavior appear when the pressure profile called for it?
 - Did the response avoid a target behavior when the pressure profile did not call for it?
 

docs/experiments/cold-wake-story-lab.md

@@ -0,0 +1,112 @@
+# Cold Wake Story Lab
+
+This experiment uses Ghostlight as a writing crutch on purpose.
+
+The goal is to write a small Cold Wake story while producing reusable projection
+training data. The story should have a protagonist, themes, locations, an arc,
+and enough room for characters to act badly for reasons the state can explain.
+
+This is not a benchmark in a clean little lab coat. It is a knife test.
+
+## Story Premise
+
+During the Cold Wake Panic, a quiet-running vessel enters a Jovian convoy
+corridor under contradictory thermal classification. The PSC tribunal freezes
+its corridor status. A Navigator route steward pushes to keep rescue live. An
+Aya-aligned sanctuary coordinator sees a possible personhood claim in a corrupted
+packet. A Cognitum shell-vendor liaison tries to keep an implicated routing
+artifact framed as derivative firmware instead of testimony.
+
+The protagonist is **Maer Tidecall**, a Cetacean Navigator route steward whose
+core problem is not whether to be kind. Maer's problem is that rescue obligation
+has to survive contact with institutions that can punish ambiguity faster than
+they can understand it.
+
+## Themes
+
+- rescue obligation versus category law
+- personhood trying not to get filed as firmware
+- care as material capacity, not decorative mercy
+- institutional fear disguised as neutral procedure
+- language as a battlefield where a person can be erased without anyone raising
+  their voice
+
+## Arc
+
+### 1. Freeze
+
+Location: Ganymede relay tribunal chamber.
+
+Maer challenges PSC adjudicator Isdra Vel after she issues a conditional risk
+hold on the vessel. The scene should establish the central wound: the corridor
+needs categories, but categories can turn rescue delay into lawful abandonment.
+
+### 2. Intake
+
+Location: Aya-aligned clinic behind Ganymede dockside pumpworks.
+
+Maer asks Sella Ren to hold sanctuary capacity for a rescue that may never be
+authorized. Sella refuses to let care become fantasy. The story should force
+Maer to see that rescue language can push costs onto care workers unless it names
+the material burden.
+
+### 3. Provenance
+
+Location: Callisto spillover compliance office.
+
+Maer confronts Veyr Oss, a Cognitum shell-vendor liaison whose routing artifact
+may explain the vessel's contradictory behavior. Veyr tries to keep the artifact
+inside product language. This is the scene most likely to produce nonverbal
+action if the pressure earns it: silence, leaving, blocking a terminal, grabbing
+a wrist, or violence. The system must not assume one more clever line is always
+the answer.
+
+### 4. Threshold
+
+Location: Jovian convoy threshold as the vessel's heat-debt window closes.
+
+Maer returns to Isdra with enough suspicion to complicate the freeze but not
+enough proof to make the room clean. The ending should not be a total victory.
+The viable resolution is a narrow, institutionally usable rescue contact that
+preserves evidence and life without pretending the category system has healed.
+
+## Experiment Loop
+
+For each scene:
+
+1. Define character-local state and scene pressure.
+2. Project the next response prompt for one character.
+3. Send only `prompt_text` to the Qwen response model.
+4. Save the model output.
+5. Review whether the output served the story, respected speaker-local context,
+   and chose speech/action/silence plausibly.
+6. Tweak the projection pipeline if the output fails in a way the pipeline
+   should have prevented.
+7. Save accepted and rejected examples as training material.
+
+## Qwen Target
+
+Use the LAN Ollama endpoint from `gamecult-ops`:
+
+- base URL: `http://192.168.1.84:11434`
+- model: `qwen3.5:9b`
+- default context: `8192`
+- use request files for `curl.exe` calls so PowerShell does not eat the JSON and
+  burp out theater.
+
+## Output Folders
+
+- agent-state fixture: `examples/agent-state.cold-wake-story-lab.json`
+- projection examples: `examples/projections/cold-wake-story-lab.jsonl`
+- readable projection mirror: `examples/projections/cold-wake-story-lab.pretty.json`
+- generated response captures: `experiments/cold-wake-story-lab/`
+
+## Acceptance Criteria
+
+- The story has a visible arc, not just disconnected argument samples.
+- Projection prompts include enough setting and cultural pressure for Qwen to
+  write without seeing the lore archive.
+- At least one projected turn permits action instead of dialogue.
+- Review notes explicitly mark whether Qwen chose speech, action, silence, or a
+  mixed beat.
+- Failures are kept when they teach a useful boundary.