diff --git a/docs/architecture/projection-distillation-plan.md b/docs/architecture/projection-distillation-plan.md index 81da1ce..d1bc00b 100644 --- a/docs/architecture/projection-distillation-plan.md +++ b/docs/architecture/projection-distillation-plan.md @@ -92,9 +92,11 @@ latency, and cost, not for final prose quality. ### Aetheria-Tuned Agent Model The character agent or responder is a separate training target. It may be -adapted on Aetheria source-grounded scenes, reviewed Ghostlight action/dialogue -receipts, and negative examples so setting priors and tone live in the model -rather than in a giant prompt sack. +adapted on a corpus of strong source-grounded Aetheria responses and reviewed +Ghostlight action/dialogue receipts spread throughout the timeline. The point is +for setting priors, tone, factional preconceptions, institutional assumptions, +species affordances, and recurring social failure patterns to live in the +responder model rather than in a giant prompt sack. That model still receives only character-local context at runtime. Fine-tuning should teach it Aetheria's pressure ecology, not grant it author vision, diff --git a/docs/architecture/soft-model-training-artifacts.md b/docs/architecture/soft-model-training-artifacts.md index 86a45e5..5f97b0c 100644 --- a/docs/architecture/soft-model-training-artifacts.md +++ b/docs/architecture/soft-model-training-artifacts.md @@ -85,10 +85,16 @@ Training artifacts: - constraints violated - reviewer notes about voice, lore, embodiment, agency, and scene fit -This model should be fine-tuned or adapted on Aetheria material eventually. A -local general model can learn the world, tone, institutions, factional pressures, -species affordances, and recurring social failure patterns better than a raw -model prompted from cold storage every time. +This model should be fine-tuned or adapted on a corpus of strong Aetheria +responses spread across the timeline. The goal is not a bolt-on lore module. +The goal is for the responder to internalize the shape of the setting: what +people assume, what institutions feel like, which pressures are normal, which +metaphors appear, what bodies can do, and how the universe tends to hurt people +while insisting it is only doing math. + +That reduces lore handholding at runtime. It does not remove character-local +context. The tuned responder still needs the specific room, speaker, listener, +state, memory, affordances, and known facts for the current turn. ### Appraiser @@ -145,10 +151,7 @@ Training artifacts: This is where confidence belongs. Do not grow a second correction-resistance organ because the first one got lonely. -### Lore And Tone Adapter - -Bakes Aetheria source material into the response model's priors so it does not -need the entire vault in every prompt. +### Aetheria Responder Corpus Training artifacts: @@ -159,11 +162,15 @@ Training artifacts: - negative examples where generic sci-fi assumptions were rejected - reviewer notes tied to AetheriaLore paths -This is the likely home for an Aetheria-tuned model built from a Qwen-class base -or another local model. The tuned model should know the smell of the setting: -Dominion severity, Aya-coded care pressure, Navigator embodiment, corporate -certification culture, insurer cruelty, upload personhood fights, factional -misreads, and the dry little horrors of systems pretending to be neutral. +This corpus trains the character agent/responder directly. It is how a +Qwen-class base or another local model learns Aetheria's native assumptions +rather than waiting for every prompt to drag the whole lore vault behind it in a +little wagon. + +The tuned responder should learn the smell of the setting: Dominion severity, +Aya-coded care pressure, Navigator embodiment, corporate certification culture, +insurer cruelty, upload personhood fights, factional misreads, and the dry +little horrors of systems pretending to be neutral. It still must receive character-local context at runtime. Fine-tuning gives it priors, not omniscience. @@ -224,6 +231,7 @@ JSONL wearing a crown. - state-mutation corpus - relationship/perception corpus - lore-grounding digest corpus +- Aetheria responder corpus - Ink branch annotation corpus - institution/faction/consumer decision corpus - negative-example corpus diff --git a/notes/fresh-workspace-handoff.md b/notes/fresh-workspace-handoff.md index 89e24d4..89369b7 100644 --- a/notes/fresh-workspace-handoff.md +++ b/notes/fresh-workspace-handoff.md @@ -128,8 +128,8 @@ first reviewed mutation receipt showing how a selected generated branch changes both involved characters. The soft-model training doctrine is now explicit: deterministic gates stay code, while fuzzy judgments become reviewed artifacts for future projector, character-agent, appraiser, mutator, -relationship/perception, lore-tone, and institution/faction/consumer decision -models. +relationship/perception, Aetheria responder, and institution/faction/consumer +decision models. ## Current Direction @@ -315,10 +315,11 @@ Remaining projection path: - train a smaller student projector only after the artifact schema, input slicer, and evaluator stabilize - treat the character agent/responder as a training target too: an - Aetheria-tuned local model can eventually learn lore, tone, factional priors, - species affordances, and institutional pressure from source-grounded - Aetheria corpora plus reviewed Ghostlight receipts, while still acting only - inside character-local context and deterministic mutation gates + Aetheria-tuned local model can eventually learn lore, tone, factional + preconceptions, species affordances, and institutional pressure from a corpus + of strong responses spread throughout the timeline plus reviewed Ghostlight + receipts, while still acting only inside character-local context and + deterministic mutation gates ## Warnings diff --git a/notes/ghostlight-current-system-map.md b/notes/ghostlight-current-system-map.md index b41729e..696de47 100644 --- a/notes/ghostlight-current-system-map.md +++ b/notes/ghostlight-current-system-map.md @@ -80,7 +80,7 @@ state and re-entry discipline. - doctrine for treating every non-deterministic judgment as reviewed training material for future specialized models, including projector, character agent/responder, appraiser, state mutator, relationship/perception - classifier, lore-tone adapter, and institution/faction/consumer decision + classifier, Aetheria responder, and institution/faction/consumer decision models - `docs/architecture/aetheria-cold-wake-training-fixture.md` - human-facing note that reframes the recovered Cold Wake scenario as @@ -280,7 +280,9 @@ retries improved Maer's Navigator embodiment but exposed object-custody drift, invalid appraisal paths, Qwen tool dropout, and prompt-constraint leakage into responder prose. The soft-model artifact doctrine now makes the broader training boundary explicit: deterministic gates stay code, while fuzzy -projection, action choice, appraisal, mutation, relationship movement, lore-tone -adaptation, and economic/faction decision judgments become reviewed artifacts -for later fine-tuning or distillation. Do not materialize another capture until -it has no failure or repair notes. +projection, action choice, appraisal, mutation, relationship movement, +Aetheria responder training, and economic/faction decision judgments become +reviewed artifacts for later fine-tuning or distillation. The lore/tone goal is +not a separate adapter; it is a responder trained on strong timeline-spread +Aetheria responses until the setting's assumptions become native priors. Do not +materialize another capture until it has no failure or repair notes. diff --git a/notes/ghostlight-implementation-plan.md b/notes/ghostlight-implementation-plan.md index 28c96f9..a675aa6 100644 --- a/notes/ghostlight-implementation-plan.md +++ b/notes/ghostlight-implementation-plan.md @@ -221,17 +221,18 @@ - Train or adapt a smaller student projector only after the artifact schema, input slicer, and evaluator are stable. - Treat the character agent/responder as a training target too. A - Qwen-class or similar local model can eventually be adapted on the - Aetheria corpus and reviewed Ghostlight receipts so lore, tone, factional - priors, species affordances, and institutional pressure are baked into the - response model's priors rather than pasted into every prompt. + Qwen-class or similar local model can eventually be adapted on a corpus of + strong Aetheria responses spread throughout the timeline, plus reviewed + Ghostlight receipts, so lore, tone, factional preconceptions, species + affordances, and institutional pressure become native responder priors + rather than prompt handholding. - Keep deterministic gates as code even after fine-tuning: visibility, action legality, object custody, resource accounting, schema validation, source provenance, mutation authority, prompt leakage checks, and Ink compilation. - Emit reviewed artifacts for every fuzzy organ before training specialized models: projector, character agent/responder, participant appraiser, - state mutator, relationship/perception classifier, lore-tone adapter, and + state mutator, relationship/perception classifier, Aetheria responder, and later institution/faction/consumer decision models. 6. Build the first drama-scaffolding loop. diff --git a/state/evidence.jsonl b/state/evidence.jsonl index 0de2a6b..bc7b01a 100644 --- a/state/evidence.jsonl +++ b/state/evidence.jsonl @@ -47,3 +47,4 @@ {"ts": "2026-05-02T03:29:51+00:00", "type": "documentation", "status": "ok", "note": "Added source-first Aetheria generation doctrine and patched AetheriaLore Cetacean Navigators with habitat/interface ergonomics so Navigator action affordances can be projected from source rather than local improvisation."} {"ts": "2026-05-02T03:46:05+00:00", "type": "experiment", "status": "mixed", "note": "Source-checked Navigator projection improved Maer action proposals: v13-v17 route packets through shared clinic displays and wet/dry interface affordances instead of human body defaults. No capture was materialized; failures exposed object-custody drift, invented mutation paths, Qwen appraisal tool dropouts, and prompt-constraint leakage into Sella prose."} {"ts": "2026-05-02T07:55:40+00:00", "type": "architecture", "status": "ok", "note": "Soft judgments that cannot be deterministic code are now explicit training-artifact targets: projector, character agent/responder, participant appraiser, state mutator, relationship/perception classifier, lore-tone adapter, and later institution/faction/consumer decision models should all emit reviewed accepted/rejected receipts before fine-tuning or distillation."} +{"ts": "2026-05-02T07:58:01+00:00", "type": "architecture", "status": "ok", "note": "Clarified Aetheria lore/tone training: the responder itself should be fine-tuned on strong source-grounded responses across the timeline so setting assumptions become native priors, rather than treating lore/tone as a separate adapter layer."} diff --git a/state/map.yaml b/state/map.yaml index cd4d4f1..a408f06 100644 --- a/state/map.yaml +++ b/state/map.yaml @@ -137,9 +137,9 @@ accepted_design: - Ghostlight's Aetheria generation doctrine is source-first and source-repairing: if a blocking lore gap appears while generating scenes or training data, patch AetheriaLore with a narrow elaboration before treating the detail as available to Ghostlight projection. - AetheriaLore now includes Navigator habitat and interface ergonomics for mixed wet/dry, acoustic, translation, and shared operational spaces, so future Navigator scenes should cite that source instead of improvising body affordances. - Source-checked Navigator retries v11-v17 improved Maer's generated action affordances but were not materialized; remaining blockers are object-custody semantics, valid appraisal/mutation paths, Qwen tool reliability, and prompt-constraint leakage into responder prose. - - The soft organs are training targets: projector, character agent/responder, appraiser, state mutator, relationship/perception classifier, lore-tone adapter, and later institution/faction/consumer decision models should all produce reviewed artifacts for future distillation or fine-tuning. - - An Aetheria-tuned agent model may bake in lore, tone, factional priors, species affordances, and institutional pressure, but it still acts only inside Ghostlight's character-local context, deterministic gates, and mutation authority. + - The soft organs are training targets: projector, character agent/responder, appraiser, state mutator, relationship/perception classifier, Aetheria responder, and later institution/faction/consumer decision models should all produce reviewed artifacts for future distillation or fine-tuning. + - An Aetheria-tuned responder should learn lore, tone, factional preconceptions, species affordances, and institutional pressure from strong source-grounded responses spread throughout the timeline, but it still acts only inside Ghostlight's character-local context, deterministic gates, and mutation authority. 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, a complete Cold Wake short story assembled from accepted projection outputs, a hand-authored Ink-backed Cold Wake branching scene, the first Qwen-generated Ink draft materialized from local awareness with reviewed training annotations, the first reviewed sequential branch mutation that updates both Maer and Sella after a sanctuary intake beat, a v2 sequential Qwen capture proving the symmetrical turn shape produces better action behavior, a v5 sequential Qwen capture proving local thinking-plus-tools invocation works for the scene runner, the first projected-local-context seam that routes sequential prompts through character-local operating prose instead of raw state variables, tightened receipt handling for projector-routed Qwen failures, a default no-thinking policy for strict sequential tool calls, a v10 no-thinking projector-routed capture accepted as draft, the first materialized projector-routed sequential Ink scene with reviewed mutation training data, a source-checked embodiment/interface projection surface for Navigator scenes, a formal source-first Aetheria generation doctrine, v11-v17 source-checked retry receipts showing improved Maer embodiment plus remaining Sella response/prompt-leak blockers, and a soft-model training artifact doctrine for turning every non-deterministic judgment into future projector, agent, appraiser, mutator, relationship, lore-tone, or decision-model training data. + 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, a complete Cold Wake short story assembled from accepted projection outputs, a hand-authored Ink-backed Cold Wake branching scene, the first Qwen-generated Ink draft materialized from local awareness with reviewed training annotations, the first reviewed sequential branch mutation that updates both Maer and Sella after a sanctuary intake beat, a v2 sequential Qwen capture proving the symmetrical turn shape produces better action behavior, a v5 sequential Qwen capture proving local thinking-plus-tools invocation works for the scene runner, the first projected-local-context seam that routes sequential prompts through character-local operating prose instead of raw state variables, tightened receipt handling for projector-routed Qwen failures, a default no-thinking policy for strict sequential tool calls, a v10 no-thinking projector-routed capture accepted as draft, the first materialized projector-routed sequential Ink scene with reviewed mutation training data, a source-checked embodiment/interface projection surface for Navigator scenes, a formal source-first Aetheria generation doctrine, v11-v17 source-checked retry receipts showing improved Maer embodiment plus remaining Sella response/prompt-leak blockers, and a soft-model training artifact doctrine for turning every non-deterministic judgment into future projector, agent/responder, appraiser, mutator, relationship/perception, Aetheria responder, or decision-model training data. next_action: Revise the Sella next-action prompt/rendering path so response constraints stay out of in-world prose, then rerun the source-checked Navigator sequential pass and materialize only a capture with no failure or repair notes.