Clean up architecture documentation
This commit is contained in:
Meta Vens
parent
109891c30a
commit
43eddcd9d6
11 changed files with 328 additions and 49 deletions
|
|
@ -137,15 +137,21 @@ The first useful organs are:
|
|||
3. relationship and faction stance records
|
||||
4. event and scene records
|
||||
5. prompt projection renderer for content generation
|
||||
6. drama beat generator driven by goals, values, pressure, and memory
|
||||
7. dialogue scaffold renderer that preserves stance, mask, and relationship
|
||||
6. coordinator/story runtime driven by goals, values, pressure, memory,
|
||||
continuity, branch flags, and author constraints
|
||||
7. branch compiler that turns reviewed branch plans into Ink, sidecars,
|
||||
callbacks, consequence variables, and visual prompt handles
|
||||
8. IF artifact reviewer that catches decorative choices, fake folds, unused
|
||||
variables, and missing consequence callbacks
|
||||
9. dialogue scaffold renderer that preserves stance, mask, and relationship
|
||||
asymmetry
|
||||
8. character-local dialogue context packs that restrict each generated voice to
|
||||
10. character-local dialogue context packs that restrict each generated voice to
|
||||
what that character knows, feels, wants, fears, and perceives
|
||||
|
||||
The output should be inspectable. The author should be able to see why the
|
||||
machine thinks a character would deflect, confess, threaten, soften, bargain,
|
||||
or lie.
|
||||
lie, leave, spend a resource, conceal evidence, or make a choice that changes a
|
||||
later scene.
|
||||
|
||||
## What This Is Not
|
||||
|
||||
|
|
@ -166,10 +172,12 @@ The first implementation is useful if it can:
|
|||
|
||||
- keep character state distinct from prompt prose
|
||||
- render dialogue from character-local context rather than global plot vibes
|
||||
- generate dialogue scaffolds that reflect relationship history
|
||||
- generate branching dialogue/action scaffolds that reflect relationship history
|
||||
- create procedural drama from incompatible goals instead of random hostility
|
||||
- preserve cultural and factional pressure without making clones
|
||||
- produce story beats a human author can actually use
|
||||
- make branch choices matter through later affordances, callbacks, visual
|
||||
modifiers, costs, risks, and endings
|
||||
- explain its choices in state terms rather than vibes in a rented coat
|
||||
|
||||
The real test is not whether the system can improvise a cool scene once. The
|
||||
|
|
|
|||
|
|
@ -62,7 +62,7 @@ One-shot generation of both sides is allowed only as bootstrap scaffolding, and
|
|||
must be marked as such in the capture review. It is not the target runtime
|
||||
shape.
|
||||
|
||||
The first generated draft path is:
|
||||
Archived local-model draft path:
|
||||
|
||||
```text
|
||||
examples/agent-state.cold-wake-story-lab.json
|
||||
|
|
@ -74,11 +74,11 @@ examples/agent-state.cold-wake-story-lab.json
|
|||
-> examples/ink/cold-wake-sanctuary-intake.qwen-draft.training.json
|
||||
```
|
||||
|
||||
The Qwen draft is accepted as training material, not as final polished prose or
|
||||
runtime architecture. It generated Maer's branch options and Sella's responses
|
||||
in one pass, which is a known shortcut. A generated branch can still be useful
|
||||
because it exposes a decision shape, an action type, a misread, or a prompt
|
||||
failure even when the line writing still needs a human with a knife.
|
||||
This path is a receipt from the old local-model materialization experiment. It
|
||||
is useful for validation and negative examples, not the current gold-data path.
|
||||
The current target path is source-grounded coordinator state, projected local
|
||||
context, reviewed character turns, branch compiler materialization, and IF
|
||||
artifact review.
|
||||
|
||||
## Ink Comments
|
||||
|
||||
|
|
@ -139,12 +139,15 @@ Manual reviewed until trained:
|
|||
- activation changes
|
||||
- promotion from branch outcome into durable Aetheria lore
|
||||
|
||||
The first prototype lives in
|
||||
The first early prototype lives in
|
||||
`examples/ink/cold-wake-sanctuary-intake.ink`.
|
||||
|
||||
The first Qwen-generated draft lives in
|
||||
The archived Qwen-generated draft lives in
|
||||
`examples/ink/cold-wake-sanctuary-intake.qwen-draft.ink`.
|
||||
|
||||
The current accepted-as-draft branch-and-fold scaffold lives in
|
||||
`examples/ink/cold-wake-branch-and-fold.v0.ink`.
|
||||
|
||||
## Cross-Scene Consequence Carryover
|
||||
|
||||
Ghostlight branching scenes should not behave like disconnected dialogue samples.
|
||||
|
|
|
|||
|
|
@ -97,7 +97,7 @@ The sequential runtime still owns:
|
|||
- event proposal capture
|
||||
- participant appraisal
|
||||
- reviewed mutation
|
||||
- Ink materialization
|
||||
- branch compiler / Ink materialization
|
||||
|
||||
The projector owns the translation from Ghostlight's internal state language to
|
||||
character-local operating prose.
|
||||
|
|
@ -106,7 +106,7 @@ character-local operating prose.
|
|||
|
||||
The first projector is deterministic and deliberately simple. It uses threshold
|
||||
language like "dominant", "strong", "present", and "background" rather than
|
||||
numeric scores. This is enough to remove raw state soup from Qwen prompts and
|
||||
numeric scores. This is enough to remove raw state soup from responder prompts and
|
||||
create inspectable training artifacts.
|
||||
|
||||
Later versions can become smarter, but they should still preserve this boundary:
|
||||
|
|
|
|||
|
|
@ -1,7 +1,16 @@
|
|||
# Qwen Invocation Notes
|
||||
|
||||
Ghostlight should call the local Qwen box through Ollama's chat/tool path when
|
||||
structured output matters.
|
||||
These notes are archived local-model plumbing guidance. They are useful when a
|
||||
task explicitly tests the neighboring Qwen box, Ollama tool calls, or old
|
||||
capture validation. They are not the current gold-data generation path.
|
||||
|
||||
For accepted responder, branch compiler, or IF review data, preserve exact
|
||||
visible inputs, raw outputs, review labels, and coordinator interventions. Do
|
||||
not treat old local-model captures as active steering unless the task is
|
||||
explicitly model plumbing.
|
||||
|
||||
When Ghostlight does call the local Qwen box, use Ollama's chat/tool path for
|
||||
structured output.
|
||||
|
||||
## Verified Local Path
|
||||
|
||||
|
|
@ -71,7 +80,7 @@ Sources:
|
|||
|
||||
## Current Policy
|
||||
|
||||
For structured Ghostlight generation:
|
||||
For archived local-model experiments and future explicit Qwen plumbing:
|
||||
|
||||
1. Prefer native chat/tool calls with thinking disabled for strict generation.
|
||||
2. Keep tool schemas narrow and concrete.
|
||||
|
|
|
|||
|
|
@ -360,28 +360,35 @@ agent state + scene state
|
|||
|
||||
## First Prototype
|
||||
|
||||
The next implementation should not try to simulate a city. Calm down, wizard.
|
||||
The first implementation should not try to simulate a city. Calm down, wizard.
|
||||
|
||||
Prototype one tiny Ink-backed branching scene loop:
|
||||
Prototype one small Ink-backed branching scene loop:
|
||||
|
||||
1. Load `examples/agent-state.cold-wake-story-lab.json`.
|
||||
2. Select one scene and one acting agent.
|
||||
3. Build a local awareness packet from existing state.
|
||||
4. Compile projection controls.
|
||||
5. Generate or author an Ink scene with branch choices and NPC responses.
|
||||
6. Compile the Ink scene to JSON with `inkjs`.
|
||||
7. Save a sidecar training annotation that maps Ink branches to Ghostlight
|
||||
5. Generate or author branch choices and NPC responses from local context.
|
||||
6. Materialize the branch plan through the Branch Compiler as Ink, sidecar
|
||||
annotation, compiler notes, consequence variables, and visual prompt handles.
|
||||
7. Compile the Ink scene to JSON with `inkjs`.
|
||||
8. Save a sidecar training annotation that maps Ink branches to Ghostlight
|
||||
state basis, actor intent, consequence surfaces, and mutation policy.
|
||||
8. Apply selected branch consequences only through reviewed mutation:
|
||||
9. Apply selected branch consequences only through reviewed mutation:
|
||||
- Ink variables can track local playthrough facts
|
||||
- canonical state changes require resolver/reviewer approval
|
||||
9. Manually review fuzzy participant appraisals and state effects each turn:
|
||||
10. Run IF artifact review before acceptance:
|
||||
- every material variable is read later or marked telemetry-only
|
||||
- every material choice changes affordance, risk, callback, visual state, or
|
||||
outcome
|
||||
- folds preserve consequence rather than hiding it
|
||||
11. Manually review fuzzy participant appraisals and state effects each turn:
|
||||
- update every affected participant before selecting the next actor
|
||||
- add one memory or belief update
|
||||
- adjust one or two `current_activation` values
|
||||
- update one relationship stance if justified
|
||||
- record the rationale as training data
|
||||
10. Validate the resulting fixture.
|
||||
12. Validate the resulting fixture.
|
||||
|
||||
The goal is not literary quality or autonomous plot generation. The goal is to
|
||||
prove that a character can drive a scene beat, produce branchable dialogue or
|
||||
|
|
|
|||
|
|
@ -9,7 +9,7 @@ Some parts should be boring code:
|
|||
- action legality
|
||||
- object custody
|
||||
- resource accounting
|
||||
- Ink materialization
|
||||
- Ink syntax validation and compilation
|
||||
- event append
|
||||
- provenance checks
|
||||
- deterministic context slicing
|
||||
|
|
@ -295,6 +295,66 @@ little horrors of systems pretending to be neutral.
|
|||
It still must receive character-local context at runtime. Fine-tuning gives it
|
||||
priors, not omniscience.
|
||||
|
||||
### Branch Compiler
|
||||
|
||||
Turns scene intent, reviewed branch candidates, consequence packets, fold plans,
|
||||
and visual continuity requirements into playable Ink plus sidecar receipts.
|
||||
|
||||
The Branch Compiler is trainable because good branching structure is not just
|
||||
syntax. It needs to decide where choices belong, which consequences can fold
|
||||
into compact variables, which choices deserve route splits, which callbacks must
|
||||
exist later, and how visual modifiers attach to actual state. Deterministic code
|
||||
can validate Ink and refs. It cannot, by itself, know whether a branch feels
|
||||
meaningful or whether convergence is lying.
|
||||
|
||||
Training artifacts:
|
||||
|
||||
- fixture brief
|
||||
- source-grounded lore digest refs
|
||||
- input branch candidates
|
||||
- reviewed consequence packets
|
||||
- fold plan
|
||||
- generated Ink
|
||||
- generated sidecar
|
||||
- compiler notes
|
||||
- variable set/read map
|
||||
- route split or fold rationale
|
||||
- visual prompt handles
|
||||
- reviewer findings and repair labels
|
||||
|
||||
The compiler does not own hidden character truth, participant appraisal, or
|
||||
durable social mutation. It materializes structure from already reviewed or
|
||||
explicitly marked draft material.
|
||||
|
||||
### IF Artifact Reviewer
|
||||
|
||||
Audits playable branch artifacts before they enter the accepted corpus.
|
||||
|
||||
This reviewer is separate from the branch compiler because compilers are very
|
||||
good at producing valid little lies. The reviewer checks whether tracked state
|
||||
actually matters, whether choices change future affordances or outcomes, whether
|
||||
folds preserve callbacks, whether visual modifiers correspond to real branch
|
||||
state, and whether endings read the variables the story told us were important.
|
||||
|
||||
Training artifacts:
|
||||
|
||||
- Ink fixture under review
|
||||
- sidecar under review
|
||||
- coordinator artifact
|
||||
- compiler notes
|
||||
- validation output
|
||||
- accepted/rejected decision
|
||||
- severity-ranked findings
|
||||
- failure labels such as `potemkin_state`, `cosmetic_choice`, `fake_fold`,
|
||||
`missing_state_read`, `state_named_instead_of_checked`,
|
||||
`visual_callback_missing`, and `ending_ignores_major_state`
|
||||
- required fixes
|
||||
- repaired artifact refs, when available
|
||||
|
||||
The first Cold Wake branch-and-fold review is seed material for this organ
|
||||
because it caught decorative variables before they were allowed to become
|
||||
training sludge.
|
||||
|
||||
### Institution, Faction, And Consumer Decision Models
|
||||
|
||||
Eventually Ghostlight should support economic behavior and major faction
|
||||
|
|
@ -354,6 +414,8 @@ JSONL wearing a crown.
|
|||
- lore-grounding digest corpus
|
||||
- Aetheria responder corpus
|
||||
- Ink branch annotation corpus
|
||||
- Branch compiler corpus
|
||||
- IF artifact review corpus
|
||||
- institution/faction/consumer decision corpus
|
||||
- negative-example corpus
|
||||
|
||||
|
|
|
|||
|
|
@ -160,6 +160,8 @@ world state + author constraints + lore grounding
|
|||
-> state mutator
|
||||
-> relationship/perception updater
|
||||
-> coordinator carries hooks, glue prose, and next beat
|
||||
-> branch compiler materializes branch-and-fold Ink plus sidecar receipts
|
||||
-> IF artifact reviewer accepts, rejects, or sends the fixture back
|
||||
```
|
||||
|
||||
Some stages are trained as generative models. Some are trained as classifiers or
|
||||
|
|
@ -208,6 +210,8 @@ Approximate target scale:
|
|||
| Participant appraiser | 300 appraisals | 3,000+ appraisals | 10,000+ appraisals |
|
||||
| State mutator | 500 accepted / 200 rejected | 5,000+ patches | 10,000+ patches |
|
||||
| Relationship/perception updater | 500 updates | 5,000+ updates | 20,000+ sequence updates |
|
||||
| Branch compiler | 25 fixtures / 25 rejected | 500 fixtures | 2,000+ fixtures |
|
||||
| IF artifact reviewer | 100 reviewed fixtures / 100 labeled failures | 2,000+ labeled fixtures | 10,000+ labeled fixtures |
|
||||
| Guardrail/evaluator classifiers | 300 labeled failures | 3,000+ labeled outputs | 20,000+ labeled outputs |
|
||||
| Institution/faction/consumer decisions | 500 decisions | 5,000+ decisions | 20,000+ decisions |
|
||||
| Technology/item manifests | 100 records / 50 rejected | 1,000+ manifests | 5,000+ manifests |
|
||||
|
|
@ -237,6 +241,8 @@ Candidate stages:
|
|||
- coordinator/story runtime
|
||||
- projector, if deterministic projection becomes too brittle
|
||||
- character agent/responder
|
||||
- branch compiler, if rule-assisted Ink materialization is too narrow for
|
||||
authored fixture variety
|
||||
- state mutator, only after the mutation schema is very stable
|
||||
- institution/faction planner, if it needs explanations and mixed outputs
|
||||
- item or technology manifest generator, if structured extraction from
|
||||
|
|
@ -258,9 +264,11 @@ Candidate stages:
|
|||
- participant appraiser
|
||||
- relationship/perception updater
|
||||
- output evaluator
|
||||
- IF artifact reviewer
|
||||
- candidate action ranker
|
||||
- lore/source violation detector
|
||||
- prompt-leak detector
|
||||
- branch-consequence and fake-fold detector
|
||||
- technology and item manifest evaluator
|
||||
|
||||
### Embedding Or Retrieval Model
|
||||
|
|
@ -843,7 +851,127 @@ Pilot schema shakedown gate:
|
|||
- repeated-contact sequences, not only isolated events
|
||||
- evaluation includes asymmetric relationships and stubborn misreads
|
||||
|
||||
### 10. Output Evaluator And Guardrail Classifiers
|
||||
### 10. Branch Compiler
|
||||
|
||||
Purpose: turn reviewed scene intent, branch candidates, consequence packets, and
|
||||
fold plans into playable Ink plus sidecar receipts.
|
||||
|
||||
Current implementation:
|
||||
|
||||
- coordinator-authored Ink fixtures
|
||||
- documented branch-and-fold contract
|
||||
- manual materialization and review
|
||||
|
||||
Inputs:
|
||||
|
||||
- fixture brief and scene spine
|
||||
- source-grounded lore digest
|
||||
- current world, scene, resource, relationship, and branch state
|
||||
- actor-local branch candidates, including speech and non-speech actions
|
||||
- responder outputs or coordinator-authored placeholder reactions
|
||||
- reviewed consequence packets
|
||||
- branch-and-fold plan
|
||||
- visual continuity requirements
|
||||
- schema and Ink formatting constraints
|
||||
|
||||
Outputs:
|
||||
|
||||
- readable `.ink` scene content
|
||||
- Ink variables, choices, knots, gathers, and conditionals
|
||||
- `.training.json` sidecar tying branches to state basis, action intent,
|
||||
consequences, callbacks, and review status
|
||||
- compiler notes for fold decisions, route-split decisions, and known review
|
||||
risks
|
||||
- base image prompt handles and branch/state modification handles
|
||||
|
||||
Training architecture:
|
||||
|
||||
- start as reviewed structured authoring
|
||||
- generative decoder LLM for fixture drafting once sidecar shape stabilizes
|
||||
- deterministic validators for Ink syntax, branch ids, sidecar refs, and schema
|
||||
shape
|
||||
|
||||
Training artifacts:
|
||||
|
||||
- accepted and rejected compiled fixtures
|
||||
- Ink/sidecar pairs
|
||||
- compiler notes
|
||||
- fold and route-split decisions
|
||||
- state-variable read/write maps
|
||||
- visual prompt modifier maps
|
||||
- reviewer findings and repair receipts
|
||||
|
||||
Pilot schema shakedown gate:
|
||||
|
||||
- 25 accepted fixtures across multiple story shapes
|
||||
- 25 rejected or needs-revision fixtures with labeled compiler failures
|
||||
- every accepted fixture has an Ink file, sidecar, compiler notes, and review
|
||||
status
|
||||
- every material variable is read later or marked telemetry-only before
|
||||
acceptance
|
||||
- branch compiler output is reviewed by the IF artifact reviewer before it
|
||||
enters the accepted corpus
|
||||
|
||||
### 11. IF Artifact Reviewer
|
||||
|
||||
Purpose: catch false consequence before branch fixtures become training data.
|
||||
|
||||
Current implementation:
|
||||
|
||||
- manual/frontier-model review
|
||||
- failure taxonomy documented in
|
||||
`docs/architecture/ink-branching-scenes.md`
|
||||
|
||||
Inputs:
|
||||
|
||||
- Ink file
|
||||
- `.training.json` sidecar
|
||||
- coordinator artifact
|
||||
- branch compiler notes
|
||||
- declared variables, resources, risks, relationship bands, and visual
|
||||
modifiers
|
||||
- intended branch-and-fold plan
|
||||
- validation output
|
||||
|
||||
Outputs:
|
||||
|
||||
- `accepted`, `needs_revision`, or `rejected`
|
||||
- severity-ranked findings
|
||||
- fake-variable audit
|
||||
- choice-consequence audit
|
||||
- fold/callback audit
|
||||
- state-gated prose audit
|
||||
- visual-continuity audit
|
||||
- required fixes
|
||||
|
||||
Training architecture:
|
||||
|
||||
- classifier or cross-encoder for known failure labels
|
||||
- LLM judge for nuanced IF design failures, with reviewed labels saved
|
||||
- deterministic support tools for variable set/read maps and branch-id
|
||||
consistency
|
||||
|
||||
Training artifacts:
|
||||
|
||||
- accepted fixtures
|
||||
- rejected fixtures
|
||||
- labeled failures such as `potemkin_state`, `cosmetic_choice`, `fake_fold`,
|
||||
`missing_state_read`, `state_named_instead_of_checked`,
|
||||
`visual_callback_missing`, and `ending_ignores_major_state`
|
||||
- reviewer rationales
|
||||
- repaired fixture diffs tied to findings
|
||||
|
||||
Pilot schema shakedown gate:
|
||||
|
||||
- 100 reviewed fixtures or fixture fragments
|
||||
- 100 labeled failures, including at least 25 fake-variable or fake-fold cases
|
||||
- reviewer catches variables that do not affect affordances, risk, appraisal,
|
||||
callbacks, visual state, or outcomes
|
||||
- reviewer catches prose that names alternate branch states instead of checking
|
||||
variables
|
||||
- reviewer catches endings that ignore major prior state
|
||||
|
||||
### 12. Output Evaluator And Guardrail Classifiers
|
||||
|
||||
Purpose: reject bad model outputs before they become artifacts or state.
|
||||
|
||||
|
|
@ -888,7 +1016,7 @@ Pilot schema shakedown gate:
|
|||
- 300 labeled failures across prompt leakage, omniscience, wrong body, wrong
|
||||
lore, invalid mutation, schema drift, and flattened character agency
|
||||
|
||||
### 11. Institution, Faction, And Consumer Decision Models
|
||||
### 13. Institution, Faction, And Consumer Decision Models
|
||||
|
||||
Purpose: model non-dialogue decisions in economic, factional, institutional, and
|
||||
open-world contexts.
|
||||
|
|
@ -957,7 +1085,7 @@ Pilot schema shakedown gate:
|
|||
- include examples where factions differ in starting pre-Elysium tech base,
|
||||
manufacturing rights, maintenance ability, and supply-chain dependency
|
||||
|
||||
### 12. Technology And Item Manifest Generator
|
||||
### 14. Technology And Item Manifest Generator
|
||||
|
||||
Purpose: turn worldbuilding exploration into game-usable item, component,
|
||||
assembly, technology, and supply-chain records.
|
||||
|
|
@ -1061,19 +1189,23 @@ Recommended order:
|
|||
1. Keep deterministic validators and artifact schemas ahead of data volume.
|
||||
2. Build coordinator artifact schema and start saving coordinator receipts.
|
||||
3. Expand projected local context and character-turn capture receipts.
|
||||
4. Build evaluator labels from existing failures.
|
||||
5. Fine-tune or LoRA the Aetheria responder first, because strong response data
|
||||
4. Stabilize branch compiler artifacts for Ink, sidecars, fold notes, and visual
|
||||
prompt handles.
|
||||
5. Build IF artifact reviewer labels from existing branching failures.
|
||||
6. Build broader evaluator labels from responder, projector, lore, and mutation
|
||||
failures.
|
||||
7. Fine-tune or LoRA the Aetheria responder first, because strong response data
|
||||
directly improves visible output and reduces lore handholding.
|
||||
6. Add future-branch post-Rupture examples before claiming the responder can
|
||||
8. Add future-branch post-Rupture examples before claiming the responder can
|
||||
handle Elysium-era concepts without heavy prompting.
|
||||
7. Train projector only if deterministic projection becomes a bottleneck.
|
||||
8. Train appraiser and relationship updater after enough manual mutation receipts
|
||||
9. Train projector only if deterministic projection becomes a bottleneck.
|
||||
10. Train appraiser and relationship updater after enough manual mutation receipts
|
||||
exist.
|
||||
9. Train state mutator last among social organs because bad mutation corrupts
|
||||
11. Train state mutator last among social organs because bad mutation corrupts
|
||||
the world state.
|
||||
10. Build item manifest schema before large future-branch exploration produces
|
||||
12. Build item manifest schema before large future-branch exploration produces
|
||||
technology data.
|
||||
11. Train institution/faction/consumer models after scene-local decisions provide
|
||||
13. Train institution/faction/consumer models after scene-local decisions provide
|
||||
enough grounded examples.
|
||||
|
||||
## Minimum Corpus Before First Experimental Fine-Tune
|
||||
|
|
@ -1110,6 +1242,23 @@ A first coordinator fine-tune wants:
|
|||
- world-state refs in every example
|
||||
- glue prose paired with structured state deltas or explicit no-delta notes
|
||||
|
||||
A first branch compiler model wants:
|
||||
|
||||
- 25 accepted Ink fixtures with sidecars and compiler notes
|
||||
- 25 rejected or needs-revision fixtures with specific compiler failure labels
|
||||
- variable set/read maps for every fixture
|
||||
- branch-and-fold decisions marked as fold, conditional variant, or route split
|
||||
- visual prompt handles tied to actual state variables
|
||||
- IF artifact reviewer decisions for every accepted example
|
||||
|
||||
A first IF artifact reviewer wants:
|
||||
|
||||
- 100 reviewed fixtures or fixture fragments
|
||||
- 100 labeled failures
|
||||
- examples where the human review found fake state or fake folds before repair
|
||||
- repaired versions paired with the original finding
|
||||
- held-out fixtures with intentionally subtle cosmetic-choice and callback gaps
|
||||
|
||||
A first appraiser/classifier training run wants:
|
||||
|
||||
- 300 participant appraisals
|
||||
|
|
|
|||
|
|
@ -27,7 +27,7 @@ behavior, major faction decision-making, and technology/item manifest generation
|
|||
## Current Focus
|
||||
|
||||
The active path is not the archived local-model prototype runner. The active
|
||||
path is:
|
||||
path is a source-grounded branching-scene data loop:
|
||||
|
||||
1. Source-ground Aetheria context.
|
||||
2. Project canonical state into character-local operating context.
|
||||
|
|
@ -37,6 +37,11 @@ path is:
|
|||
leakage audit, and coordinator interventions.
|
||||
6. Apply only reviewed mutations into state, memory, relationships, perceived
|
||||
overlays, and unresolved hooks.
|
||||
7. Let the coordinator carry continuity, unresolved hooks, glue prose, and the
|
||||
next scene/beat plan.
|
||||
8. Let the branch compiler materialize playable Ink plus sidecar and compiler
|
||||
notes from reviewed state and branch decisions.
|
||||
9. Run IF artifact review before accepting the fixture as training data.
|
||||
|
||||
## Live Seams
|
||||
|
||||
|
|
@ -73,6 +78,11 @@ path is:
|
|||
- Research sanity probe: `experiments/research-sanity/navigator-embodiment-sanity-v0.capture.json`
|
||||
- Packet builder: `tools/build_responder_packet.py`
|
||||
- Packet/output validators: `tools/validate_responder_packets.py`, `tools/validate_responder_outputs.py`
|
||||
- Clean Cold Wake IF scaffold: `examples/ink/cold-wake-branch-and-fold.v0.ink`
|
||||
- Clean Cold Wake IF sidecar: `examples/ink/cold-wake-branch-and-fold.v0.training.json`
|
||||
- Clean Cold Wake coordinator artifact: `examples/coordinator/cold-wake-branch-and-fold.v0.json`
|
||||
- Clean Cold Wake readable run: `experiments/cold-wake-story-lab/cold-wake-branch-and-fold-clean-run.v0.md`
|
||||
- Branch compiler and IF reviewer contract: `docs/architecture/ink-branching-scenes.md`
|
||||
|
||||
Archived prototype captures and tools remain as evidence under `experiments/`,
|
||||
`examples/ink/`, and old `run_qwen_*` helpers. They are useful receipts, not the
|
||||
|
|
@ -130,6 +140,11 @@ IF scaffold with visual prompt anchors and reviewed consequence surfaces.
|
|||
- Keep persistent state small. Git history and artifacts own chronology.
|
||||
- Do not let notes, map, handoff, and evidence become four versions of the same brain.
|
||||
- Do not train on coordinator-repaired responder prose unless the repair is labeled.
|
||||
- Do not accept branch fixtures without branch compiler notes and IF artifact
|
||||
review once that seam is available.
|
||||
- Do not let tracked variables, authority pressure, heat, evidence quality,
|
||||
relationship bands, or visual modifiers remain decorative. If they are
|
||||
tracked, they need consequence or an explicit telemetry-only label.
|
||||
- Do not let Elysium branch futures overwrite Sol single-history facts.
|
||||
- Do not rebuild the older corridor-crisis game slice by accident.
|
||||
- Do not treat parent-visible subagent output as proof of research. If the
|
||||
|
|
|
|||
|
|
@ -33,9 +33,17 @@ Aetheria lore/source context
|
|||
-> sandboxed responder output
|
||||
-> review and leakage audit
|
||||
-> mutation receipt
|
||||
-> updated fixture / future training corpus
|
||||
-> updated scene/world/social state
|
||||
-> coordinator continuity and next-beat plan
|
||||
-> branch compiler materializes Ink + sidecar + compiler notes
|
||||
-> IF artifact reviewer audits consequence, fold, and visual continuity
|
||||
-> accepted fixture / future training corpus
|
||||
```
|
||||
|
||||
Responder packets are one seam, not the whole project. The current branching
|
||||
scene path also requires coordinator receipts, branch compiler artifacts, and an
|
||||
independent IF review pass before a fixture is accepted.
|
||||
|
||||
## Important Contracts
|
||||
|
||||
- Agent state: `schemas/agent-state.schema.json`
|
||||
|
|
@ -45,6 +53,8 @@ Aetheria lore/source context
|
|||
- Coordinator artifacts: `schemas/coordinator-artifact.schema.json`
|
||||
- Responder packets: `schemas/responder-packet.schema.json`
|
||||
- Responder outputs: `schemas/responder-output.schema.json`
|
||||
- Ink branch contract: `docs/architecture/ink-branching-scenes.md`
|
||||
- Training stages and corpus gates: `docs/architecture/training-plan.md`
|
||||
|
||||
## Current Live Example
|
||||
|
||||
|
|
@ -90,6 +100,17 @@ Aetheria lore/source context
|
|||
- Parent sessions currently receive only the subagent final message, not a
|
||||
visible tool-call transcript. Treat subagent research as self-reported unless
|
||||
the worker returns explicit research notes or a future runner captures calls.
|
||||
- Clean accepted-as-draft IF scaffold:
|
||||
`examples/ink/cold-wake-branch-and-fold.v0.ink`
|
||||
- Its sidecar is
|
||||
`examples/ink/cold-wake-branch-and-fold.v0.training.json`.
|
||||
- Its coordinator artifact is
|
||||
`examples/coordinator/cold-wake-branch-and-fold.v0.json`.
|
||||
- Its readable run is
|
||||
`experiments/cold-wake-story-lab/cold-wake-branch-and-fold-clean-run.v0.md`.
|
||||
- This fixture is coordinator/IF scaffold data. It is not raw responder gold.
|
||||
- The user review that exposed fake variables and cosmetic state is now seed
|
||||
taxonomy for the IF artifact reviewer.
|
||||
|
||||
## Archived Receipts
|
||||
|
||||
|
|
@ -99,7 +120,7 @@ embodiment, prompt leakage, and materialization. They are not active direction.
|
|||
Do not let them steer gold responder data unless the task is explicitly model
|
||||
plumbing.
|
||||
|
||||
## Missing Organs
|
||||
## Missing Or Incomplete Organs
|
||||
|
||||
- Deterministic speaker-local input slicer
|
||||
- Full prompt renderer
|
||||
|
|
@ -107,13 +128,16 @@ plumbing.
|
|||
- Relationship/perception updater
|
||||
- State mutator model
|
||||
- Student projector
|
||||
- Branch compiler implementation beyond coordinator-authored fixtures
|
||||
- IF artifact reviewer implementation beyond manual/frontier review
|
||||
- Full scene/event loop implementation
|
||||
- Culture prior engine
|
||||
- Automatic promotion of branch outcomes into canonical state
|
||||
|
||||
## Current North Star
|
||||
|
||||
Generate high-quality, sandboxed, source-grounded responder/coordinator samples
|
||||
that can later train specialized Ghostlight models while staying usable by a
|
||||
game engine: exact inputs, exact outputs, provenance, review labels, state
|
||||
mutation receipts, and clear branch consequences.
|
||||
Generate high-quality, sandboxed, source-grounded branching-scene samples that
|
||||
can later train specialized Ghostlight models while staying usable by a game
|
||||
engine: exact inputs, exact outputs, provenance, review labels, state mutation
|
||||
receipts, branch compiler artifacts, IF review findings, and clear branch
|
||||
consequences.
|
||||
|
|
|
|||
|
|
@ -34,3 +34,4 @@
|
|||
{"ts": "2026-05-02T22:15:47+00:00", "type": "review", "status": "ok", "note": "Revised Cold Wake IF scaffold again after user review: authority_pressure now creates concrete PSC audit/sanction consequences, can make strong evidence unsafe without provisional_category, and appears in visual/climax modifiers."}
|
||||
{"ts": "2026-05-02T22:29:00+00:00", "type": "design", "status": "ok", "note": "IF artifact reviewer required before acceptance to catch Potemkin state variables, cosmetic choices, fake folds, missing state reads, state-named-instead-of-checked prose, visual callback gaps, and endings that ignore major state."}
|
||||
{"ts": "2026-05-02T22:30:27+00:00", "type": "design", "status": "ok", "note": "Branch Compiler documented as a separate organ: it materializes playable Ink, sidecars, fold/callback structure, and visual handles from fixture state and reviewed consequence packets, without owning hidden character truth or durable social mutation."}
|
||||
{"ts": "2026-05-02T22:40:03+00:00", "type": "documentation", "status": "ok", "note": "Documentation cleanup aligned training plan, system map, handoff, soft-artifact doctrine, Ink architecture, Qwen notes, and content targets with current branch compiler + IF reviewer pipeline; old Qwen/local-model paths are archived plumbing, not active gold-data steering."}
|
||||
|
|
|
|||
|
|
@ -14,9 +14,9 @@ mission:
|
|||
social consequences, and eventually storylines, consumer behavior,
|
||||
faction decisions, and technology/item manifests for Aetheria.
|
||||
near_term: >
|
||||
Produce sandboxed, source-grounded responder training data through
|
||||
projected local context, responder packets, reviewed outputs, and audited
|
||||
mutation receipts.
|
||||
Produce sandboxed, source-grounded branching-scene training data through
|
||||
projected local context, responder packets, reviewed outputs, mutation
|
||||
receipts, coordinator continuity, branch compiler artifacts, and IF review.
|
||||
voidbot_goal: >
|
||||
Keep the appraiser, relationship/perception updater, and memory machinery
|
||||
useful for VoidBot's core companion mission: reading specific humans over
|
||||
|
|
@ -81,7 +81,7 @@ live_architecture:
|
|||
|
||||
active_subgoals:
|
||||
- Keep Ghostlight focused on the architecture and data pipeline for socially persistent agents.
|
||||
- Build schema-stable, sandboxed responder data before chasing larger runtime loops.
|
||||
- Build schema-stable, sandboxed responder, coordinator, branch compiler, and IF review data before chasing larger runtime loops.
|
||||
- Preserve Aetheria grounding through source-backed lore, branch labels, and reviewed assumptions.
|
||||
- Shape every soft judgment as future training data, not permanent prompt folklore.
|
||||
- Keep item/technology exploration database-shaped for Aetheria-Economy instead of prose-only worldbuilding, including colony-consumed manufactured trade goods, not only gear assemblies.
|
||||
|
|
@ -154,6 +154,7 @@ current_artifacts:
|
|||
- experiments/cold-wake-story-lab/cold-wake-branch-and-fold-clean-run.v0.md
|
||||
archived_receipts:
|
||||
- Archived local-model prototype captures remain under experiments/ and examples/ink/ for validation and lessons, not active steering.
|
||||
- docs/architecture/qwen-invocation-notes.md is archived local-model plumbing guidance, not the current gold-data route.
|
||||
|
||||
data_flow:
|
||||
- Rehydrate from map, handoff, system map, implementation plan, and `npm run state:status`.
|
||||
|
|
|
|||
Loading…
Reference in a new issue