From 723f1a40a4cb3056c4bf75f5b93f2eab94ea3bef Mon Sep 17 00:00:00 2001 From: Meta Vens Date: Fri, 1 May 2026 23:36:55 +0100 Subject: [PATCH] Preserve open-world training ambitions --- docs/architecture/sequential-agent-runtime.md | 45 ++++++++++++++----- notes/fresh-workspace-handoff.md | 9 +++- notes/ghostlight-current-system-map.md | 3 +- notes/ghostlight-implementation-plan.md | 13 ++++-- state/evidence.jsonl | 1 + state/map.yaml | 6 ++- 6 files changed, 59 insertions(+), 18 deletions(-) diff --git a/docs/architecture/sequential-agent-runtime.md b/docs/architecture/sequential-agent-runtime.md index ef0a8b1..8880050 100644 --- a/docs/architecture/sequential-agent-runtime.md +++ b/docs/architecture/sequential-agent-runtime.md @@ -4,10 +4,14 @@ Ghostlight's near-term target is not a general-purpose life simulator. The near-term product target is procedural interactive dialogue: set up a scene, let characters choose believable responses, expose player-facing choices, and -propagate consequences into state, memory, and social perception. Later, the -same machinery can help generate longer storylines, but that is not license to -build an everything-simulator right now. The little goblin will ask for a crown. -Do not give it one. +propagate consequences into state, memory, and social perception. + +The long-term target is broader. The same classifiers, appraisers, and +projection models should eventually operate inside a simulated game world with +open-world contexts, iterated decisions, economic behavior, and major-faction +decision-making. Do not train a tiny dialogue-only creature that panics the +first time someone buys medicine, defects from a faction, or reacts to a +shortage. Also do not build the whole economy today. Both cliffs are stupid. The runtime target for now is scene-local: choose which character is acting, give that character local awareness, let them act, apply consequences, mutate @@ -42,7 +46,10 @@ For game use, the loop should be able to emit: - unresolved hooks for later scenes If a feature does not help produce those artifacts, it belongs in the parking -lot until the smaller machine works. +lot until the smaller machine works. If a field or label preserves information +needed for later open-world behavior, consumer choice, faction pressure, or +iterated decisions, keep it visible in the training artifact even if the first +prototype does not simulate it. ## Local Awareness @@ -195,7 +202,8 @@ already learned human interpretation from twelve examples and a strong coffee. ## Scope Boundary -Build only what interactive dialogue generation needs first. +Build interactive dialogue generation first, but shape the data for later +open-world use. In scope now: @@ -206,18 +214,35 @@ In scope now: - manual reviewed appraisals - memory, belief, relationship, and activation deltas - branch consequences that can be carried into later scenes +- action intent and listener interpretation labels +- resource, price, scarcity, obligation, reputation, and faction-pressure tags + when they affect the decision +- reviewed examples of consumer-like choices and institution/faction pressure, + even when the prototype only records them as scene state -Out of scope for now: +Training-data obligations now: + +- Preserve why an agent chose, bought, refused, hoarded, traded, complied, + defected, concealed, punished, or helped. +- Preserve what the agent believed the cost, risk, status effect, and social + consequence would be. +- Preserve faction and institution pressures as perceived by the agent, not + only as objective setting lore. +- Preserve iterated-decision context: promises, debts, prior betrayals, + reputation, memory, scarcity, and expected future contact. + +Out of implementation scope for now: - full autonomous world simulation -- economy simulation +- economy simulation loops - city-scale scheduling - long-horizon plot invention without author scaffolding -- autonomous factions moving offscreen +- autonomous factions moving offscreen as full actors - physical simulation beyond scene affordances - automatic social appraisal rules pretending to be trained classifiers -The eventual storyline generator should grow out of reliable scene machinery, +The eventual storyline generator, economic behavior model, and faction decision +model should grow out of reliable scene machinery and reviewed decision data, not the other way around. ## Projection's Narrower Job diff --git a/notes/fresh-workspace-handoff.md b/notes/fresh-workspace-handoff.md index 88b39e5..cc36a29 100644 --- a/notes/fresh-workspace-handoff.md +++ b/notes/fresh-workspace-handoff.md @@ -141,8 +141,13 @@ Remaining projection path: awareness, produce branchable dialogue or action, and mutate only the state it has authority to affect - keep the near-term product target to interactive dialogue trees and - scene-local consequences; defer world-scale simulation and autonomous - long-horizon plot machinery + scene-local consequences +- preserve training-data hooks for future open-world behavior, iterated + decisions, psychologically grounded consumer choices, scarcity/reputation + pressure, and major faction or institution decision-making +- defer implementing world-scale simulation loops, autonomous offscreen + factions, economy loops, and long-horizon plot machinery until the scene-local + machine works - use a frontier teacher model to generate/audit projection artifacts - train a smaller student projector only after the artifact schema, input slicer, and evaluator stabilize diff --git a/notes/ghostlight-current-system-map.md b/notes/ghostlight-current-system-map.md index d68eb2b..0191e8e 100644 --- a/notes/ghostlight-current-system-map.md +++ b/notes/ghostlight-current-system-map.md @@ -48,7 +48,8 @@ state and re-entry discipline. - `docs/architecture/sequential-agent-runtime.md` - scene-local dialogue/runtime loop where agents act from local awareness, generated actions become event proposals, and bounded state mutation drives - branchable dialogue consequences + branchable dialogue consequences while preserving training hooks for later + open-world, economic, and faction decision models - `docs/experiments/cold-wake-story-lab.md` - four-scene Cold Wake writing experiment using Ghostlight projection as a story-generation crutch and training-data source diff --git a/notes/ghostlight-implementation-plan.md b/notes/ghostlight-implementation-plan.md index e8ccbcc..0de1558 100644 --- a/notes/ghostlight-implementation-plan.md +++ b/notes/ghostlight-implementation-plan.md @@ -41,6 +41,9 @@ - Keep the product target narrow: procedural interactive dialogue trees for games first, then scene-local character action with consequence propagation, then longer procedural storyline generation later. + - Do not make the training data dialogue-only. Preserve labels and state + hooks for open-world contexts, iterated decisions, consumer behavior, + scarcity, reputation, faction pressure, and institution decision-making. - Build local awareness from scene state, agent state, relationships, memories, resources, affordances, and known constraints. - Keep action primitives concrete: speech, silence, movement, gesture, @@ -62,9 +65,13 @@ - Emit game-useful artifacts: scene transcript, candidate player choices, NPC response branches, state/memory/social-perception deltas, and unresolved hooks. - - Keep out of scope for now: world-scale simulation, economy, city-scale - scheduling, autonomous offscreen factions, and long-horizon plot invention - without author scaffolding. + - Keep out of implementation scope for now: world-scale simulation loops, + economy simulation loops, city-scale scheduling, autonomous offscreen + factions as full actors, and long-horizon plot invention without author + scaffolding. + - Still record why agents buy, refuse, trade, hoard, comply, defect, conceal, + punish, or help when those decisions appear in scenes; those reviewed + decisions are seed data for later consumer-behavior and faction models. - First prototype: rerun one Cold Wake story-lab beat through a deterministic local-awareness/projection-control renderer, then apply one conservative state mutation and validate the resulting fixture. diff --git a/state/evidence.jsonl b/state/evidence.jsonl index 0ef866f..058c51b 100644 --- a/state/evidence.jsonl +++ b/state/evidence.jsonl @@ -31,3 +31,4 @@ {"ts": "2026-05-01T21:59:49+00:00", "type": "design", "status": "accepted", "note": "Reframed Ghostlight target from specific dialogue projection toward sequential character agents: each acting agent gets local awareness, options, objectives, and mutation authority; generated action becomes an event proposal resolved into bounded state changes. Single-turn projection remains a debug/training seam, not the procedural-generation endpoint."} {"ts": "2026-05-01T22:18:59+00:00", "type": "design", "status": "accepted", "note": "Sequential runtime prototype should keep fuzzy interpretation manual at first. Action tools should be concrete primitives; ask/refuse/offer/reveal/conceal/request are intended communicative functions on actions, while listener appraisal may misread them. Manual appraisal and mutation review becomes training data for future classifiers."} {"ts": "2026-05-01T22:31:04+00:00", "type": "design", "status": "accepted", "note": "Tightened Ghostlight runtime scope: near-term goal is procedural interactive dialogue trees and scene-local consequence propagation into state, memory, and social perception. Sequential agent machinery should serve branchable game dialogue first; world-scale simulation, autonomous offscreen factions, economy, and long-horizon plot generation stay deferred."} +{"ts": "2026-05-01T22:36:38+00:00", "type": "design", "status": "accepted", "note": "Adjusted scope boundary: implement scene-local interactive dialogue first, but shape training data for later open-world agents, iterated decisions, psychologically grounded consumer behavior, scarcity/reputation pressure, and major faction or institution decision-making. Defer full simulation loops, not the labels and state hooks those future models will need."} diff --git a/state/map.yaml b/state/map.yaml index bc0e6dd..acec023 100644 --- a/state/map.yaml +++ b/state/map.yaml @@ -12,7 +12,8 @@ objective: Turn Ghostlight into a legible machine for persistent socially aware characters that can generate interactive dialogue trees for games, propagate scene consequences into state and memory, and eventually support procedural - storyline generation without collapsing into prompt soup. + storyline generation, psychologically grounded consumer behavior, and major + faction decision-making without collapsing into prompt soup. current: > Move from hand-authored single-turn projection artifacts toward the first scene-local sequential dialogue tick: local awareness, action or dialogue @@ -103,7 +104,8 @@ accepted_design: - Generated action should become an event proposal resolved by mutation authority, not direct unrestricted edits to world or other-agent private state. - Sequential runtime action primitives should stay concrete; ask, refuse, offer, reveal, conceal, and request are communicative intentions carried by actions, while listener interpretation remains separate and may be wrong. - Early fuzzy social updates should be manual reviewed training data, not brittle rule-based appraisals pretending to understand relationships before the corpus exists. - - Scope boundary: generate interactive dialogue trees and scene-local consequences first; defer world-scale simulation, autonomous offscreen factions, economy, and long-horizon plot invention until the smaller machine works. + - Scope boundary: implement interactive dialogue trees and scene-local consequences first, but preserve training data hooks for open-world contexts, iterated decisions, economic behavior, scarcity, reputation, faction pressure, and institution decision-making. + - Do not train dialogue-only models that will be useless for later game-world simulation; reviewed scene data should preserve why agents buy, refuse, trade, hoard, comply, defect, conceal, punish, or help when those decisions appear. current_status: summary: Ghostlight now has the Epiphany-style persistence spine, the first architecture payload, the v0 canonical agent-state schema seam, prompt projection docs, immediate schema fixes from evaluation, first reviewed Cat/Oz projection examples, a revised Aetheria lore-grounding architecture, a lore grounding digest schema, the first draft Cold Wake historical grounding digest, a Cold Wake story-lab fixture, projection controls, validated Qwen capture receipts, and a complete Cold Wake short story assembled from accepted projection outputs.