diff --git a/notes/fresh-workspace-handoff.md b/notes/fresh-workspace-handoff.md index 7f05a47..3c11fe1 100644 --- a/notes/fresh-workspace-handoff.md +++ b/notes/fresh-workspace-handoff.md @@ -45,7 +45,6 @@ Ghostlight now has the persistence spine plus the first architecture payload: - Aetheria lore grounding docs: - `docs/architecture/aetheria-lore-grounding-architecture.md` - `docs/architecture/lore-grounding-digest-format.md` - - `docs/architecture/qwen-invocation-notes.md` - `docs/architecture/projected-local-context.md` - `docs/architecture/soft-model-training-artifacts.md` - `docs/architecture/training-plan.md` @@ -81,48 +80,17 @@ Ghostlight now has the persistence spine plus the first architecture payload: - `docs/experiments/cold-wake-story-lab.md` - `examples/agent-state.cold-wake-story-lab.json` - `examples/projections/cold-wake-story-lab.jsonl` - - `examples/projections/cold-wake-story-lab.pretty.json` - `experiments/cold-wake-story-lab/the-narrowest-possible-margin.md` - - `experiments/cold-wake-story-lab/scene-03-maer-veyr.qwen3-5-9b.capture.json` - - `experiments/cold-wake-story-lab/scene-03-maer-veyr.qwen3-5-9b.v2.capture.json` - - `experiments/cold-wake-story-lab/scene-03-maer-veyr.qwen3-5-9b.v3.capture.json` - - additional scene 1, 2, 3, and 4 Qwen captures under - `experiments/cold-wake-story-lab/` + - archived local-model prototype captures under `experiments/cold-wake-story-lab/` - `tools/build_cold_wake_story_lab_fixture.py` - - `tools/run_qwen_projection.py` - - `tools/validate_qwen_captures.py` -- Ink branching seam: +- Ink branching and mutation seam: - `examples/ink/cold-wake-sanctuary-intake.ink` - `examples/ink/cold-wake-sanctuary-intake.training.json` - - `examples/ink/cold-wake-sanctuary-intake.qwen-draft.ink` - - `examples/ink/cold-wake-sanctuary-intake.qwen-draft.training.json` - - `experiments/ink/cold-wake-sanctuary-intake-qwen-branch-candidates-v1.capture.json` - - `experiments/ink/cold-wake-sanctuary-intake-qwen-branch-candidates-v1.prompt.md` - - `experiments/ink/cold-wake-sanctuary-intake-qwen-branch-candidates-v1.response.md` - - `tools/run_qwen_ink_branch_generation.py` - - `tools/materialize_qwen_ink_draft.py` + - archived prototype Ink/capture receipts under `examples/ink/` and `experiments/ink/` - `tools/validate_ink_examples.py` - - `tools/run_qwen_ink_sequential_generation.py` - `tools/materialize_sequential_capture.py` - `tools/apply_sequential_ink_branch_mutation.py` - `inkjs` dev dependency and `npm run ink:validate` -- sequential branch mutation seam: - - `experiments/ink/cold-wake-sanctuary-intake-qwen-sequential-v1.capture.json` - - `experiments/ink/cold-wake-sanctuary-intake-qwen-sequential-v2.capture.json` - - `experiments/ink/cold-wake-sanctuary-intake-qwen-sequential-v3.capture.json` - - `experiments/ink/cold-wake-sanctuary-intake-qwen-sequential-v5.capture.json` - - `experiments/ink/cold-wake-sanctuary-intake-qwen-sequential-v6-projector.capture.json` - - `experiments/ink/cold-wake-sanctuary-intake-qwen-sequential-v8-projector-tightened.capture.json` - - `experiments/ink/cold-wake-sanctuary-intake-qwen-sequential-v9-projector-tightened.capture.json` - - `experiments/ink/cold-wake-sanctuary-intake-qwen-sequential-v10-no-think.capture.json` - - `experiments/ink/cold-wake-sanctuary-intake-qwen-sequential-v1.mutation.json` - - `experiments/ink/qwen-chat-tools-smoke.json` - - `examples/agent-state.cold-wake-story-lab.after-sanctuary-ledger.json` -- projector-routed sequential materialization seam: - - `examples/ink/cold-wake-sanctuary-intake.qwen-sequential-v10.ink` - - `examples/ink/cold-wake-sanctuary-intake.qwen-sequential-v10.training.json` - - `examples/agent-state.cold-wake-story-lab.after-v10-packet-assessment.json` - - `experiments/ink/cold-wake-sanctuary-intake-qwen-sequential-v10.mutation.json` - projected local context seam: - `schemas/projected-local-context.schema.json` - `examples/projected-contexts/scene-02-sanctuary-intake.maer_tidecall.projected-context.json` @@ -150,7 +118,7 @@ Ghostlight now has the persistence spine plus the first architecture payload: There is still no full runtime yet. The live machine is design, state discipline, a cleaner repo boundary, the first executable state contract, a documented projection path, the first Ink-backed playable branching scene, the -first Qwen-generated Ink draft with reviewed training annotations, and the +first prototype-generated Ink draft with reviewed training annotations, and the 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 @@ -270,8 +238,8 @@ Run a sandboxed no-fork research-enabled responder against allowing only the declared AetheriaLore scope. Capture exact visible input, consulted refs, research summary, raw output, review labels, and leakage audit. Review whether the output uses institutional lore and latent pressure without -prompt parroting, omniscience, or mission-critical trauma-dumping. Do not call -Qwen for gold responder data yet. +prompt parroting, omniscience, or mission-critical trauma-dumping. Do not use +archived local-model prototype runners for gold responder data. Cat/Oz remains useful as an Elysium procedural mechanics fixture, but single-history grounding should start from authored historical lore rather than @@ -333,112 +301,12 @@ The evaluation follow-up fixed the immediate v0 schema issues: Completed projection path items: -- define `schemas/projection-example.schema.json` -- create reviewed Cat/Oz projection examples -- choose Cold Wake Panic as the first authored historical flashpoint -- fill the first draft lore grounding digest: - `examples/lore-grounding/cold-wake-panic.ganymede-corridor.json` -- define the Cold Wake story lab as a four-scene arc with Maer Tidecall as - protagonist -- add the first Cold Wake projection example and first Qwen response capture -- widen projection from dialogue-only turns to response turns that can include - action, silence, withdrawal, violence, or mixed beats -- revise the first Maer/Veyr projection through v3 so unresolved - personhood/claimant suspicion stays unresolved in Qwen output -- save both useful failed captures and the accepted v3 response as training - material -- complete the first readable Cold Wake story, - `experiments/cold-wake-story-lab/the-narrowest-possible-margin.md`, with a - receipts table tying accepted story beats to projection and capture artifacts -- promote projection controls into the projection artifact shape: - `frame_controls`, `authority_boundaries`, `object_custody`, - `required_semantics`, and `forbidden_resolutions` -- add a Qwen projection runner and capture validator so response artifacts are - reviewed training receipts, not loose logs -- adopt Ink as the standard playable branching scene format rather than - inventing a parallel Ghostlight dialogue-tree format -- add the first Cold Wake Ink scene, - `examples/ink/cold-wake-sanctuary-intake.ink`, with a reviewed sidecar - annotation tying branches to Ghostlight state, projection controls, - consequence surfaces, and manual mutation policy -- add the first local-awareness-to-Qwen-to-Ink pass: - `tools/run_qwen_ink_branch_generation.py` produced a reviewed capture, and - `tools/materialize_qwen_ink_draft.py` materialized - `examples/ink/cold-wake-sanctuary-intake.qwen-draft.ink` -- add the first sequential Qwen branch capture: - `experiments/ink/cold-wake-sanctuary-intake-qwen-sequential-v1.capture.json` - is useful-needs-revision because it separated actor and responder context but - lacked the per-turn participant appraisal boundary and returned non-canonical - action labels -- add the first reviewed state mutation replay: - `tools/apply_sequential_ink_branch_mutation.py` writes - `experiments/ink/cold-wake-sanctuary-intake-qwen-sequential-v1.mutation.json` - and - `examples/agent-state.cold-wake-story-lab.after-sanctuary-ledger.json`, - updating Maer and Sella state, relationship stance, memories, perceived - overlays, and the event log after the selected event -- add the second sequential Qwen capture: - `experiments/ink/cold-wake-sanctuary-intake-qwen-sequential-v2.capture.json` - uses the symmetrical turn framing and produces better action behavior - (`show_object` followed by Sella's `withhold_object`), but remains - useful-needs-revision because prose prompting did not force array-typed JSON - fields -- verify Qwen invocation: - `tools/check_qwen_chat_tools.py` proves local `qwen3.5:9b` returns both - `message.thinking` and native `tool_calls` through Ollama `/api/chat` -- update the sequential runner to use `/api/chat` with native tools and - strict tool calls -- add v3 and v5 thinking-plus-tools captures: - v3 preserved a useful failure where old prompt text still caused plain JSON - content in later passes; v5 is accepted as draft and produced valid - tool-shaped Maer choice, Sella appraisal, and Sella next action -- identify the projector gap: - the sequential runner is still scaffolding because it sends selected raw - numeric state variables into prompts. The next real organ is a projector or - state interpreter that turns canonical variables, memories, relationship - stance, culture, and scene pressure into readable character-local operating - context and action affordances. Character agents should not need to interpret - `current_activation`, `plasticity`, or numeric dimension scores. -- add the first projector seam: - `tools/project_local_context.py` creates - `ghostlight.projected_local_context.v0` artifacts and rendered prompt prose - from canonical state, scene, relationship, memories, projection controls, and - event context while hiding raw state variables from the prompt text -- add projected-context validation: - `tools/validate_projected_contexts.py` rejects projected prompt text that - leaks raw state markers such as `current_activation`, `plasticity`, means, or - decimal state values -- route the sequential Qwen runner through projected local context: - `tools/run_qwen_ink_sequential_generation.py` now sends projected operating - prose instead of selected activation dictionaries -- add v6 projector-routed Qwen capture: - `experiments/ink/cold-wake-sanctuary-intake-qwen-sequential-v6-projector.capture.json` - validates the architecture boundary and produces usable Maer/Sella behavior, - but remains useful-needs-revision because Qwen stringified a nested - `response_constraints` field in Sella appraisal -- tighten the sequential runner: - repair notes are separated from fatal validation notes, obvious truncated - stringified tool arrays can be repaired, failed Maer choice generation writes - a reviewed capture instead of exiting, and Sella's next-action prompt receives - appraisal context as prose constraints rather than raw numeric deltas -- add v8 and v9 projector-routed failure receipts: - v8 shows malformed stringified `choices`; v9 shows a no-tool-call dropout. - These are Qwen tool reliability failures, so no projector-routed branch was - materialized into Ink in this pass. -- disable thinking by default for strict sequential generation: - the v9 thinking trace showed Qwen looping on schema self-checks for minutes - before returning no tool call, so `tools/run_qwen_ink_sequential_generation.py` - now defaults to `think: false`; use `--think` only for diagnostics -- add v10 no-thinking projector-routed capture: - `experiments/ink/cold-wake-sanctuary-intake-qwen-sequential-v10-no-think.capture.json` - validates as accepted-as-draft with no repair notes or failure notes -- materialize v10: - `tools/materialize_sequential_capture.py` wrote a playable Ink branch, sidecar - training annotation, reviewed mutation receipt, and mutated Cold Wake fixture - from the accepted projector-routed capture -- update v10 capture provenance: - `accepted_into_ink_ref` now points at - `examples/ink/cold-wake-sanctuary-intake.qwen-sequential-v10.ink` +- defined projection examples, Cold Wake grounding, story-lab fixtures, and projection controls +- adopted Ink as the playable branching scene format +- built and validated local-awareness, projected-context, responder-packet, capture, and mutation seams +- preserved archived local-model prototype captures as reviewed receipts, not active guidance +- materialized one accepted projector-routed prototype branch into Ink, sidecar annotation, mutation receipt, and mutated fixture +- extracted the important lessons: character-local context, source-checked embodiment, object custody, per-turn consolidation, structured failure receipts, and no raw-state leakage into responder prompts Remaining projection path: @@ -446,13 +314,12 @@ Remaining projection path: - build a renderer that emits projection controls before prompt prose - review the embodiment-aware materialized v10 Ink and mutation receipt for narrative quality and state-change plausibility -- run another Qwen pass through the source-checked projector to test whether - patched Navigator habitat/interface lore improves Maer's generated action - proposals -- v11-v17 did improve Maer's source-grounded embodiment, but no capture was +- keep using sandboxed responder packets and reviewed outputs for fresh + generation passes +- source-checked retries improved Maer's grounded embodiment, but no capture was materialized because Sella appraisal/response still exposed object-custody - drift, invented mutation paths, Qwen tool dropout, or prompt-constraint - leakage + drift, invented mutation paths, structured-output dropout, or + prompt-constraint leakage - next pass should focus on separating response constraints from in-world prose before chasing another materialized Ink branch - decide whether to polish this Ink text by hand or generalize the sequential diff --git a/notes/ghostlight-current-system-map.md b/notes/ghostlight-current-system-map.md index a153506..3dcd2f1 100644 --- a/notes/ghostlight-current-system-map.md +++ b/notes/ghostlight-current-system-map.md @@ -71,9 +71,6 @@ state and re-entry discipline. - `docs/architecture/lore-grounding-digest-format.md` - human-facing format guide for cultural, factional, institutional, role, and speaker-boundary lore digests -- `docs/architecture/qwen-invocation-notes.md` - - verified local Qwen invocation policy: Ollama `/api/chat`, native tools, - thinking enabled, repair for known nested-argument double-stringify cases - `docs/architecture/projected-local-context.md` - contract for the projector seam that turns canonical state into character-local operating context before a response model sees it, @@ -186,7 +183,7 @@ state and re-entry discipline. Aetheria-native institutional texture, retrieval lessons, and a warning that character backstory must remain available as latent pressure - `experiments/cold-wake-story-lab/` - - Qwen response captures and reviews from the writing experiment + - archived local-model response captures and reviews from the writing experiment - `the-narrowest-possible-margin.md`, the first readable short story assembled from accepted projection outputs - `examples/lore-grounding/historical-flashpoint.template.json` @@ -201,73 +198,11 @@ state and re-entry discipline. - reviewed sidecar annotation tying Ink branches to Ghostlight local awareness, projection controls, branch rationale, consequence surfaces, and manual mutation policy -- `examples/ink/cold-wake-sanctuary-intake.qwen-draft.ink` - - first Qwen-generated playable Ink draft materialized from a Ghostlight - local-awareness prompt -- `examples/ink/cold-wake-sanctuary-intake.qwen-draft.training.json` - - sidecar annotation for the generated draft, accepted as training material - rather than final polished prose -- `examples/ink/cold-wake-sanctuary-intake.qwen-sequential-v10.ink` - - first playable Ink branch materialized from an accepted projector-routed - sequential Qwen capture -- `examples/ink/cold-wake-sanctuary-intake.qwen-sequential-v10.training.json` - - sidecar annotation for the v10 materialized sequential branch -- `experiments/ink/cold-wake-sanctuary-intake-qwen-branch-candidates-v1.capture.json` - - reviewed Qwen branch-generation receipt, including prompt, parsed response, - strengths, failure notes, and pipeline lessons -- `experiments/ink/cold-wake-sanctuary-intake-qwen-sequential-v1.capture.json` - - useful-needs-revision sequential Qwen receipt that separates Maer-local - choice generation from Sella-local next-action generation, while preserving - the discovered failure around missing per-turn appraisal/consolidation and - invalid action labels -- `experiments/ink/cold-wake-sanctuary-intake-qwen-sequential-v2.capture.json` - - useful-needs-revision sequential Qwen receipt using the symmetrical turn - model; it produced canonical action labels and a concrete Sella - `withhold_object` next action, but still needs stricter schema enforcement - because some array fields came back as strings or objects -- `experiments/ink/cold-wake-sanctuary-intake-qwen-sequential-v3.capture.json` - - useful-needs-revision thinking-plus-tools receipt that proved Maer could - call the tool while later passes still fell back to plain JSON content - because old prompt text conflicted with the tool path -- `experiments/ink/cold-wake-sanctuary-intake-qwen-sequential-v5.capture.json` - - accepted-as-draft thinking-plus-tools receipt using Ollama `/api/chat`, - native tools, and `think: true` for Maer choice, Sella appraisal, and Sella - next action -- `experiments/ink/cold-wake-sanctuary-intake-qwen-sequential-v6-projector.capture.json` - - useful-needs-revision projector-routed receipt proving Qwen prompts can use - character-local operating prose instead of raw state variables; still - exposed nested `response_constraints` stringification in appraisal output -- `experiments/ink/cold-wake-sanctuary-intake-qwen-sequential-v8-projector-tightened.capture.json` - - useful-needs-revision receipt proving failed Maer choice generation is now - captured instead of crashing the runner; exposed malformed stringified - `choices` -- `experiments/ink/cold-wake-sanctuary-intake-qwen-sequential-v9-projector-tightened.capture.json` - - useful-needs-revision receipt showing Qwen can still drop tool calls - entirely under projector-routed prompting -- `experiments/ink/cold-wake-sanctuary-intake-qwen-sequential-v10-no-think.capture.json` - - accepted-as-draft receipt showing projector-routed sequential generation - works cleanly when strict tool calls run with thinking disabled -- `examples/projected-contexts/` - - current rendered projected local context artifacts for Maer and Sella in - the sanctuary intake scene -- `experiments/ink/qwen-chat-tools-smoke.json` - - smoke receipt proving local `qwen3.5:9b` returns both `message.thinking` - and native `tool_calls` -- `experiments/ink/cold-wake-sanctuary-intake-qwen-sequential-v1.mutation.json` - - first reviewed branch mutation receipt, including manual participant - appraisal/consolidation, normalized action labels, applied state deltas, and - deferred unresolved facts -- `examples/agent-state.cold-wake-story-lab.after-sanctuary-ledger.json` - - mutated Cold Wake fixture after the selected sanctuary ledger branch, - updating Maer and Sella state, relationship stance, memories, perceived - overlays, event log, and active scene memories -- `examples/agent-state.cold-wake-story-lab.after-v10-packet-assessment.json` - - mutated Cold Wake fixture after the v10 packet-assessment branch, - preserving unresolved packet personhood while updating state, relationship - stance, memories, perceived overlays, event log, and active scene memories -- `experiments/ink/cold-wake-sanctuary-intake-qwen-sequential-v10.mutation.json` - - reviewed mutation receipt for v10 materialization, keeping fuzzy appraisal - and relationship movement manual and auditable +- archived local-model prototype Ink and capture receipts + - older files with `qwen` in the path remain in `examples/ink/` and + `experiments/ink/` as historical receipts; they are not the active + generation path and should not steer new gold-data work unless model + plumbing is explicitly being tested - `tools/ghostlight_state.py` - compact state inspection and evidence or branch updates - `tools/ghostlight_prepare_compaction.py` @@ -276,34 +211,18 @@ state and re-entry discipline. - dependency-free validator for the current schema fixture invariants - `tools/validate_projection_examples.py` - dependency-free validator for projection example JSONL records -- `tools/run_qwen_projection.py` - - sends a projection prompt to the LAN Qwen box and saves response/capture - artifacts without PowerShell encoding footguns -- `tools/validate_qwen_captures.py` - - validates reviewed Qwen response capture receipts +- archived local-model prototype tools + - older `run_qwen_*`, `materialize_qwen_*`, `validate_qwen_captures.py`, and + `check_qwen_chat_tools.py` helpers remain for validating historical + receipts and testing model plumbing; they are not the active gold-data path - `tools/validate_lore_grounding.py` - dependency-free validator for lore grounding digest examples - `tools/validate_ink_examples.py` - compiles Ink examples with `inkjs` and validates sidecar training - annotations plus reviewed Qwen Ink-generation captures -- `tools/run_qwen_ink_branch_generation.py` - - builds a local-awareness prompt from the Cold Wake fixture and asks Qwen - for Ink branch candidates -- `tools/materialize_qwen_ink_draft.py` - - turns a reviewed Qwen branch capture into a playable Ink draft and sidecar - annotation + annotations plus reviewed prototype Ink-generation captures - `tools/materialize_sequential_capture.py` - turns an accepted sequential capture into playable Ink, sidecar annotation, reviewed mutation receipt, and a mutated fixture -- `tools/run_qwen_ink_sequential_generation.py` - - builds a sequential Qwen prototype through Ollama `/api/chat` with native - tools and thinking disabled by default: actor-local choices, participant - appraisal/consolidation, then next-actor action from updated state; now - routes character context through the projected-local-context seam, captures - failed Maer choice calls, separates repair notes from fatal validation - notes, and keeps numeric appraisal deltas out of next-action prompt context -- `tools/check_qwen_chat_tools.py` - - fast smoke test for local Qwen thinking plus native tool calling - `tools/project_local_context.py` - deterministic first projector that compiles canonical state, scene facts, relationships, memories, projection controls, and optional observed event @@ -352,108 +271,14 @@ state and re-entry discipline. - an automatic Ink branch generator from local awareness - automatic promotion of Ink branch outcomes into canonical state -The canonical agent-state schema, projection example schema, lore grounding -digest schema, projected-local-context schema, coordinator artifact schema, -responder-packet schema, first draft Cold Wake grounding digest, and first Cold Wake story-lab Qwen -captures now exist as v0 seams. The -first complete Cold Wake story pass exists with receipts, projection controls -have been promoted, and Ink is now the standard branching scene format. The -first Qwen-generated Ink draft now exists with a reviewed capture, the first -selected branch now has a reviewed mutation replay that updates both involved -characters, v2 shows the symmetrical turn model improves action behavior, v5 -proves local thinking-plus-tools invocation works for the scene runner, and v6 -routes that runner through projected character-local operating context. v8 and -v9 show that Qwen tool-call reliability, not the projector seam, is now the -immediate blocker. The v9 thinking trace also showed schema self-check looping -instead of tool-call completion, so strict sequential generation now disables -thinking by default. v10 validated that correction with an accepted-as-draft -projector-routed no-thinking capture, which has now been materialized into Ink -plus reviewed mutation training data. The next implementation target is -fixing the Sella next-action prompt/rendering path after source-checked v11-v17 -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 -coordination, projection, action choice, appraisal, mutation, relationship -movement, Aetheria responder training, and economic/faction decision judgments -become reviewed artifacts for later fine-tuning or distillation. Gold responder -data now requires a sandbox boundary: the coordinator may be omniscient, but the -responder sees only the exact projected local packet, visible event, allowed -actions, and explicitly included source excerpts; artifacts preserve raw output, -reviewed output, hidden-context refs, leakage audit, isolation method, and -coordinator intervention labels. The responder packet seam now turns that -requirement into a concrete artifact and validator, so the next gold-data pass -can hand a worker the exact prompt surface instead of hoping nobody leaks the -coordinator's brain into the room. The first no-fork Sella responder capture now -exists and validates as accepted-as-draft with raw output preserved, parsed -output matched exactly, leakage audit clear, and no coordinator prose repair. -Responder data now has two explicit lanes: `packet_only` for runtime parity with -curated source excerpts, and `retrieval_augmented` for coordinator- or -retriever-selected scoped AetheriaLore refs that can bake setting priors into -final responses while preserving consulted refs. This is not autonomous -responder research unless the responder actually receives scoped repo access. -That explicit repo-access path is now named -`responder_scoped_repository_search`: the packet must tell the responder to -consult the declared lore docs before answering, show the allowed scope, and -give behavior-grounding research instructions. The capture must record consulted -refs and a research summary. -The first concrete research-enabled Sella packet has now been generated and -validates alongside the packet-only handoffs. -The first lane comparison is now captured: packet-only produced a clean -conditional repair-bay decision, while retrieval-augmented added heat-debt, -rescue-ledger, dockfall, and Aya sanctuary politics that should inform future -runtime retrieval. Follow-up review clarified the failure mode: injected -institutional lore should not crowd out character-local backstory from the -responder's visible pressure set, but the responder should not be rewarded for -surfacing that backstory aloud when the mission-critical exchange does not call -for it. -The runtime retrieval/projector correction is now concrete: projected contexts -carry source-backed retrieval requirements and latent pressure requirements, -and the responder packet builder copies active retrieval requirements into -packet source excerpts. The first Sella packet now receives Cold Wake heat-debt, -Navigator rescue-ledger, Aya sanctuary-capacity, and Ganymede/Lightsail route -obligation facts through that seam. -The concrete training plan now enumerates -eleven trainable stages and their likely model -families: generative decoder LLMs for coordinator, responder, and structured -soft outputs; classifiers or cross-encoders for appraisal, evaluation, and -relationship movement; embedding/retrieval models for memory and lore selection; -and deterministic code for gates. The listed per-stage counts are pilot -schema-shakedown gates, not robust training targets; review-assistant and runtime -targets are larger corpus tiers meant to avoid declaring victory after the first -small pile of sacrificial data merely shows where the schema is wrong. The -first coordinator artifact seam now validates through `npm run schema:validate` -and backfills the sanctuary packet-assessment beat as an accepted-as-draft -schema shakedown example. The coordinator is currently the authoring -agent: it glues scenes together, carries world-state refs and unresolved hooks, -decides which machinery runs, and emits glue prose. That output should stay -game-engine-shaped even while the prose lives in model imagination. 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. The timeline plan now includes generated possible futures in -Elysium as reviewed `future_branch` artifacts so models learn post-Rupture -concepts as branch-local canon indexed by lineage. It also distinguishes -branch attractors, fated events, technology-order constraints, quest injections, -and ordinary branch-local events so Elysium can support likely faction births, -authored rails, discovery order, and quest injection without flattening the -timeline into equal-probability soup. The technology/item manifest plan now -requires exploration to emit item families, variants, assemblies, -subassemblies, components, materials, faction tech-base access, prerequisites, -bottlenecks, compatibility rules, supply-chain consequences, and quest hooks for -both pre-Elysium starting tech and post-Rupture innovations, mapped onto -Aetheria-Economy's existing CultCache technological blueprint classes: -`SimpleCommodityData`, `CompoundCommodityData`, `ConsumableItemData`, `GearData`, -and other `EquippableItemData` subclasses. Runtime `ItemInstance` classes remain -simulation, inventory, cargo, provenance, branding, and world-state objects. -Recipes, assembly trees, compatibility rules, tooling, facilities, process -requirements, and supply-chain dependencies are blueprint metadata or -engine-schema gaps attached to the relevant `*Data` class until the economy code -has explicit storage for them. The player-facing gear fantasy is provenance -hunting: players chasing an ultimate item should trace exceptional assemblies, -manufacturer processes, repair history, counterfeit substitutions, supply-chain -lineage, and operating constraints through the galaxy. The technology pipeline is -therefore a content refinery: it grounds ideas in lore, decomposes them into -parts and supply chains, maps them to economy blueprint classes, defines -performance and instance variation, preserves review labels, and emits -database-shaped candidates instead of spreadsheet sludge. Do not materialize -another capture until it has no failure or repair notes. +The canonical schemas and first Cold Wake fixture seams now exist as v0 +contracts. Archived local-model prototype receipts proved useful for discovering +architecture lessons, but the active path is now sandboxed responder packets, +source-grounded local context, reviewed outputs, and audited mutation receipts. +Deterministic gates stay code; fuzzy coordination, 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 next gold-data pass should use the concrete +research-enabled Sella packet and preserve exact visible input, consulted refs, +research summary, raw output, reviewed output, hidden-context refs, leakage +audit, isolation method, and coordinator intervention labels. diff --git a/notes/ghostlight-implementation-plan.md b/notes/ghostlight-implementation-plan.md index a888fc9..7995739 100644 --- a/notes/ghostlight-implementation-plan.md +++ b/notes/ghostlight-implementation-plan.md @@ -87,11 +87,8 @@ and validate a sidecar training annotation that maps branches to Ghostlight state basis, action intent, consequence surfaces, and manual mutation policy. - - First generated draft: use `tools/run_qwen_ink_branch_generation.py` to - ask Qwen for branch candidates from local awareness, then materialize the - reviewed result with `tools/materialize_qwen_ink_draft.py`. - - First sequential draft: use `tools/run_qwen_ink_sequential_generation.py` - to separate actor-local choice, participant appraisal/consolidation, and + - Archived prototype drafts produced branch candidates from local awareness and proved the Ink materialization seam. + - First sequential prototype draft separated actor-local choice, participant appraisal/consolidation, and the next actor's move from updated state. The first saved capture is useful-needs-revision because it exposed the missing per-turn appraisal boundary and weak action enum discipline. @@ -105,17 +102,11 @@ action labels and a concrete Sella `withhold_object` next action, but still needs stricter schema or tool-call enforcement because array fields drifted into strings and objects. - - Qwen invocation correction: use Ollama `/api/chat` with native tools and - `think: true`, not `/api/generate` with prose-only JSON instructions. - Qwen's XML-shaped tool format belongs to the chat-template/tool-call - layer; Ghostlight artifacts do not need XML wrappers. - - Third sequential draft: v3 proved the tool path could work for Maer while - later passes still fell back to plain JSON because old prompt contract - text conflicted with tool calling. - - Fifth sequential draft: v5 is accepted as draft. It used thinking plus - native tools for Maer choice, Sella appraisal, and Sella next action. + - Archived local-model invocation lesson: model-specific chat/tool + formatting belongs in runner code and must not leak into Ghostlight + artifact formats. - Projector gap: the sequential runner still sends selected raw numeric - state variables into prompts. This is acceptable scaffolding for Qwen + state variables into prompts. This is archived scaffolding for local-model plumbing tests, not the final Ghostlight architecture. - Projector prototype: `tools/project_local_context.py` now turns canonical variables, memory, relationship stance, culture, and scene pressure into @@ -142,37 +133,29 @@ body types, location, and time period. If a blocking room-scale detail is missing, patch AetheriaLore narrowly before treating that detail as available to projection. - - Sixth sequential draft: v6 routed thinking-plus-tools generation through - projected local context instead of selected activation dictionaries. It - produced usable Maer/Sella behavior but remains useful-needs-revision - because Qwen stringified nested `response_constraints` in the appraisal - payload. + - A later sequential prototype routed structured generation through + projected local context instead of selected activation dictionaries and + proved the projector boundary while preserving malformed-output failures + as receipts. - Runner tightening: failed Maer choice generation now writes reviewed receipts instead of exiting, obvious malformed stringified tool arrays can be repaired, repair notes are separated from fatal validation notes, and Sella's next-action prompt receives appraisal prose instead of raw numeric deltas. - - Eighth and ninth sequential drafts: v8 captured malformed stringified - `choices`; v9 captured a no-tool-call dropout. Both validate the receipt - path and show Qwen tool reliability is now the blocker. - - Thinking policy correction: v9's thinking trace showed minutes of schema - self-check looping followed by no tool call. Strict sequential generation - now defaults to `think: false`; `--think` is reserved for diagnostics. - - Tenth sequential draft: v10 ran projector-routed generation with thinking - disabled and validated as accepted-as-draft with no repair or failure - notes. + - Later sequential prototype drafts captured malformed structured output, + no-call dropouts, and schema-looping as reviewed failure receipts. These + are runner/model reliability lessons, not active architecture targets. - V10 materialization: `tools/materialize_sequential_capture.py` materialized the accepted projector-routed capture into Ink, sidecar annotation, reviewed mutation receipt, and a mutated Cold Wake fixture. - - Next prototype: run another Qwen pass through the source-checked Navigator - projector, review whether the patched AetheriaLore ergonomics produce - cleaner action proposals, then decide whether to polish the branch or - generalize the materializer for additional accepted captures. + - Next prototype work should use sandboxed responder packets and reviewed + outputs; do not revive the archived local-model runner unless explicitly + testing model plumbing. - Source-checked retry results: v11-v17 improved Maer's body/interface behavior substantially, producing packet actions through shared clinic displays and mixed wet/dry affordances instead of humanoid action. No capture was materialized because later passes exposed repair notes, - object-custody drift, invented mutation paths, one Qwen appraisal + object-custody drift, invented mutation paths, one local-model appraisal tool-call dropout, and prompt-constraint leakage into Sella prose. - Validation correction: the sequential runner now checks Maer embodiment, Cold Wake object semantics, Sella appraisal paths/relationship ids, and @@ -252,7 +235,7 @@ - Treat projection outputs as response turns, not dialogue-only turns: characters may speak, act, withdraw, stay silent, use violence, or combine speech and action when state and scene pressure justify it. - - Keep failed and accepted Qwen captures together. The Maer/Veyr v1 and v2 + - Keep failed and accepted prototype captures together. The Maer/Veyr v1 and v2 outputs are useful failed examples; v3 is the first accepted output that preserves unresolved claimant status while allowing controlled physical action. @@ -260,9 +243,7 @@ only prompt wording problems. Projection examples now include controls for frame ownership, authority boundaries, object custody, required semantics, and forbidden resolutions. - - Keep Qwen captures as reviewed receipts. `tools/validate_qwen_captures.py` - is part of `npm run schema:validate` so unreviewed captures do not quietly - become training data. + - Keep archived local-model captures as reviewed receipts; validators keep unreviewed captures from quietly becoming training data. - Use `experiments/cold-wake-story-lab/the-narrowest-possible-margin.md` as the first end-to-end story artifact assembled from accepted response captures plus authored connective narration. @@ -272,7 +253,7 @@ - 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 a corpus of + small 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 diff --git a/state/evidence.jsonl b/state/evidence.jsonl index 8d19842..bc5e71a 100644 --- a/state/evidence.jsonl +++ b/state/evidence.jsonl @@ -69,3 +69,4 @@ {"ts": "2026-05-02T10:08:14+00:00", "type": "implementation", "status": "ok", "note": "Implemented runtime retrieval/projector requirements. Projected local contexts now emit source-backed runtime_retrieval_requirements and latent_pressure_requirements; prompt text renders retrieved lore and latent pressure handling; responder packets copy active retrieval requirements into source_excerpts so packet-only contexts receive compact Cold Wake heat-debt, Navigator rescue-ledger, Aya sanctuary-capacity, and Ganymede/Lightsail route-obligation facts without granting repo access or forcing character backstory into speech."} {"ts": "2026-05-02T10:11:38+00:00", "type": "design", "status": "accepted", "note": "Added explicit research-enabled responder instructions. `responder_scoped_repository_search` now means the responder may inspect only the declared lore scope, must consult it before answering, and must ground behavior in retrieved lore docs; packet validation requires visible research instructions and output validation requires consulted refs plus a research summary."} {"ts": "2026-05-02T10:17:55+00:00", "type": "implementation", "status": "ok", "note": "Generated the first concrete research-enabled Sella responder packet. build_responder_packet.py now supports responder_scoped_repository_search, sets repo access open only for that lane, renders Required Lore Research, and requires consulted_refs plus research_summary."} +{"ts": "2026-05-02T10:21:51+00:00", "type": "maintenance", "status": "ok", "note": "Demoted Qwen/local-model prototype material in rehydrate surfaces. Canonical map and handoff now treat those runs as archived receipts, not active gold-data direction; next action stays on sandboxed research-enabled responder packets."} diff --git a/state/map.yaml b/state/map.yaml index 6462df2..be2f84b 100644 --- a/state/map.yaml +++ b/state/map.yaml @@ -107,9 +107,9 @@ accepted_design: - Aetheria projection training has three timeline lanes: historical grounded fixtures from authored lore, transition grounded fixtures around the Rupture/displacement, and future-branch Elysium fixtures that exercise post-Rupture concepts as conditional canon. - Lore grounding digests use `ghostlight.lore_grounding_digest.v0` records to distill authored lore into cultural pressure fields, factional misread maps, role pressures, speaker knowledge boundaries, and author-only exclusions. - The Cold Wake Panic is the first chosen historical grounding fixture: a draft Ganymede corridor digest now models thermal stealth panic, insurer category failure, Navigator rescue obligation, Lightsail reliability pressure, Aya-coded care pressure, pirate porosity, Cognitum firmware denial, and Cryonix/Rossum certification culture. - - The Cold Wake story lab uses Maer Tidecall as protagonist across a four-scene arc: tribunal freeze, sanctuary intake, Callisto provenance, and convoy threshold. The first Qwen run produced a useful mixed action/dialogue beat and exposed a prompt weakness around unresolved personhood suspicion. + - The Cold Wake story lab uses Maer Tidecall as protagonist across a four-scene arc: tribunal freeze, sanctuary intake, Callisto provenance, and convoy threshold; early prototype runs exposed prompt weaknesses around unresolved personhood suspicion. - Projection controls are now first-class in projection examples: frame_controls, authority_boundaries, object_custody, required_semantics, and forbidden_resolutions capture failure lessons before prompt prose is rendered. - - Qwen response captures are validated training receipts, not loose logs; accepted, rejected, and useful-needs-revision outputs should keep review status, response kind, strengths, failures, and pipeline lessons. + - Prototype response captures are validated training receipts, not loose logs; accepted, rejected, and useful-needs-revision outputs should keep review status, response kind, strengths, failures, and pipeline lessons. - The first complete Cold Wake short story, The Narrowest Possible Margin, is assembled from Ghostlight projection receipts plus in-session connective narration. - Single-turn projection is a debug and training seam, not the final procedural-generation shape; the near-term runtime target is scene-local interactive dialogue generation with agents acting from local awareness, producing event proposals, and mutating bounded state. - Generated action should become an event proposal resolved by mutation authority, not direct unrestricted edits to world or other-agent private state. @@ -120,33 +120,28 @@ accepted_design: - Branching scene generation should propose plausible protagonist choices, including speech, silence, movement, object use, concealment, revelation, withdrawal, violence, trade, or refusal, then record NPC responses and branch consequences such as trust loss or future hooks. - Ink is the standard branching scene format for playable interactive fiction; Ghostlight should project and annotate Ink scenes rather than inventing a parallel dialogue-tree language. - Ink sidecar training annotations preserve actor-local awareness, projection controls, branch rationale, action intent, reviewed consequences, and manual mutation policy while keeping Ink itself as the branch structure. - - Qwen can generate Ink branch candidates from a Ghostlight local-awareness packet, but generated branches must be saved as reviewed captures and accepted as draft/training material before they become Ink or mutation input. + - Responder models may generate Ink branch candidates from a Ghostlight local-awareness packet, but generated branches must be saved as reviewed captures and accepted as draft/training material before they become Ink or mutation input. - Character turns must be generated from character-local awareness only: protagonist choices are generated from protagonist-local state, then responder actions are generated from responder-local state plus observable protagonist action, not protagonist private intent. - Gold responder training data must be sandboxed: the responder sees only the exact projected local packet, visible event, allowed actions, and explicitly included source excerpts; it must not inherit coordinator context, hidden state, author-only plot beats, future branch plans, or repo/lore access the trained model will not have. - Responder artifacts must preserve exact visible input, raw output, reviewed output, hidden-context refs, leakage audit, isolation method, and coordinator intervention labels; coordinator repairs are useful but cannot be laundered as raw responder behavior. - Interaction turns are symmetrical: an actor acts, the action resolves into an event, every affected participant appraises and updates state, then the next actor is selected from the newly consolidated state. - Every meaningful scene event should update each affected character's state, memory, relationship stance, or perceived overlay through a reviewed mutation receipt; Ink variables remain playthrough-local and do not become canonical truth by accident. - - The first sequential Qwen capture is useful-needs-revision training material because it separated actor and responder context but lacked the per-turn participant appraisal boundary and returned non-canonical action labels. - - The second sequential Qwen capture validates the symmetrical turn shape behaviorally by producing canonical action labels and a non-speech Sella action, but still needs revision because prose prompting did not enforce array-typed JSON fields. - - Qwen structured generation should use Ollama `/api/chat` with native tools, not `/api/generate` plus prose-only JSON instructions; Qwen's XML-shaped tool format is an internal chat-template/tool-call convention, not a reason to wrap Ghostlight artifacts in XML. - - The local Qwen tool path works with thinking enabled, but nested arguments can still double-stringify, so the runner repairs known nested fields and preserves failures as reviewed captures. - - The current sequential Qwen runner is scaffolding, not the final projection architecture: it still sends selected raw numeric state variables directly to the model. - - A projector/state interpreter must sit between canonical agent state and Qwen prompts, translating numbers, memories, relationships, culture, and scene pressure into compact character-local operating context and affordances. + - Early sequential prototype captures are useful training material because they exposed actor/responder isolation, per-turn participant appraisal, action-label discipline, tool-call/schema reliability, and the need to keep model plumbing separate from final Ghostlight architecture. + - Local-model runner code is archived scaffolding, not the active gold-data path. + - A projector/state interpreter must sit between canonical agent state and responder prompts, translating numbers, memories, relationships, culture, and scene pressure into compact character-local operating context and affordances. - Character agents should not need to interpret `current_activation`, `plasticity`, or raw dimension scores; those belong to Ghostlight's internal state model and projector training data. - The first projected local context seam now exists as `ghostlight.projected_local_context.v0`: deterministic projection turns canonical state into character-local operating prose with audit refs while hiding raw state variables from prompt text. - - The sequential Qwen runner now consumes projected local context instead of selected raw activation dictionaries. - - The first projector-routed Qwen capture validates the architecture boundary but remains useful-needs-revision because Qwen still stringified a nested appraisal list field in one pass. + - The sequential prototype runner now consumes projected local context instead of selected raw activation dictionaries. + - The first projector-routed prototype captures validated the architecture boundary while preserving malformed structured-output failures as training evidence. - Projector-routed next-action prompts should receive appraisal context as prose constraints, not raw numeric state deltas from the appraisal artifact. - - Failed sequential Qwen calls must still write reviewed receipts instead of exiting without evidence. - - v8 and v9 projector-routed captures show Qwen tool-call reliability remains the blocker: malformed stringified tool arrays and no-tool-call dropouts can occur even when the projector seam is clean. - - Thinking should be disabled by default for strict Qwen scene-generation tool calls; the v9 trace showed schema self-check looping that lasted minutes and ended with no tool call. - - The first no-thinking projector-routed sequential run, v10, validated as accepted-as-draft with no repair or failure notes. - - The accepted v10 projector-routed capture has been materialized into playable Ink, sidecar training annotation, reviewed mutation receipt, and a mutated Cold Wake fixture. + - Failed sequential model calls must still write reviewed receipts instead of exiting without evidence. + - Prototype structured-output failures show model/tool reliability is a runner concern, not a reason to distort the projector seam; malformed tool arrays, no-call dropouts, and schema-looping must become reviewed failure receipts. + - The first accepted projector-routed sequential capture has been materialized into playable Ink, sidecar training annotation, reviewed mutation receipt, and a mutated Cold Wake fixture. - Materialized sequential captures should still keep fuzzy state, relationship, memory, and perceived-overlay changes manual and audited until classifiers exist. - Embodiment and interface affordances belong in projected local context, but Aetheria faction-specific affordances must be checked against lore source before projection; current Navigator source supports fluid architecture, acoustic signaling, soft navigation light, layered water motifs, communal orientation chambers, mixed-species translation/coexistence, specialized cetacean environments, and mixed wet/dry interface ergonomics. - 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. + - Source-checked Navigator retries improved Maer's generated action affordances but were not materialized; remaining blockers are object-custody semantics, valid appraisal/mutation paths, structured-output reliability, and prompt-constraint leakage into responder prose. - The soft organs are training targets: coordinator/story runtime, 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. - The coordinator is a trainable artifact: it glues scenes together, chooses next beats and acting agents, invokes specialist machinery, carries unresolved hooks, emits reviewed glue prose, and proposes world-state changes without bypassing deterministic gates. @@ -171,5 +166,5 @@ accepted_design: - Ghostlight's technology pipeline should function as a content refinery: ground ideas in lore and mechanics, decompose them into parts and supply chains, map them to Aetheria-Economy blueprint classes, define performance and instance variation, preserve review labels, and produce database-shaped candidates instead of spreadsheet sludge. 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 coordinator artifact schema seam, the first responder packet seam, the first sandboxed responder output seam, explicit packet-only versus retrieval-augmented responder lanes, the first concrete research-enabled Sella packet, the first responder lane comparison, 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, a soft-model training artifact doctrine, a concrete training plan with pilot schema-shakedown gates separated from robust corpus targets, sandboxed responder-data requirements, an explicit future-branch lane, Elysium branch event constraints, and a technology/item manifest plan for turning nebulous named tech into blueprint-grade components, assemblies, faction tech bases, supply chains, upgrade paths, gameplay hooks, provenance-hunting player fantasy, and content-refinery outputs mapped onto Aetheria-Economy's CultCache blueprint and instance model. - next_action: Run a sandboxed no-fork research-enabled responder against `examples/responder-packets/scene-02-sanctuary-intake.sella_ren.packet.research.v0.json`, allowing only the declared AetheriaLore scope, then capture exact visible input, consulted refs, research summary, raw output, review labels, and leakage audit. Review for institutional lore use, latent pressure, prompt parroting, omniscience, and mission-critical trauma-dumping. Do not call Qwen for gold responder data yet. + 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 coordinator artifact schema seam, the first responder packet seam, the first sandboxed responder output seam, explicit packet-only versus retrieval-augmented responder lanes, the first concrete research-enabled Sella packet, the first responder lane comparison, the first draft Cold Wake historical grounding digest, a Cold Wake story-lab fixture, projection controls, validated prototype capture receipts, a complete Cold Wake short story assembled from accepted projection outputs, a hand-authored Ink-backed Cold Wake branching scene, the first prototype-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, sequential prototype receipts showing the symmetrical turn shape produces better action behavior, local-model tool-call smoke receipts archived as scaffolding, 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 structured-output failures, archived local-model invocation lessons, 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, source-checked retry receipts showing improved Maer embodiment plus remaining Sella response/prompt-leak blockers, a soft-model training artifact doctrine, a concrete training plan with pilot schema-shakedown gates separated from robust corpus targets, sandboxed responder-data requirements, an explicit future-branch lane, Elysium branch event constraints, and a technology/item manifest plan for turning nebulous named tech into blueprint-grade components, assemblies, faction tech bases, supply chains, upgrade paths, gameplay hooks, provenance-hunting player fantasy, and content-refinery outputs mapped onto Aetheria-Economy's CultCache blueprint and instance model. + next_action: Run a sandboxed no-fork research-enabled responder against `examples/responder-packets/scene-02-sanctuary-intake.sella_ren.packet.research.v0.json`, allowing only the declared AetheriaLore scope, then capture exact visible input, consulted refs, research summary, raw output, review labels, and leakage audit. Review for institutional lore use, latent pressure, prompt parroting, omniscience, and mission-critical trauma-dumping. Do not use archived local-model prototype runners for gold responder data.