Add Aetheria tonal range doctrine

This commit is contained in:
Meta Vens 2026-05-03 09:12:09 +01:00
commit ef5a2f5697
10 changed files with 110 additions and 24 deletions

View file

@ -6,7 +6,7 @@ Ghostlight is a research and implementation home for socially persistent
generative agents. The core problem is not "make an LLM sound like a person."
The core problem is building an architecture where personality, culture,
memory, masks, perception, and incentives can actually survive contact with
time, relationships, and pressure.
time, relationships, routine, tonal variation, and pressure.
This repo exists so those organs can live somewhere cleaner than a side chamber
inside another project.
@ -132,6 +132,16 @@ When the user says to prepare for imminent compaction:
- Before writing or generating Aetheria scenes, check AetheriaLore for the
factions, movements, institutions, species or body types, location, and time
period involved.
- Before drafting story prose, choose an intentional tonal mode or blend for the
scene. Aetheria should support wit-with-stakes, warmth, absurdity, domestic
texture, ritual quiet, wonder, horror, noir, dour poetry, and dry systems
prose as needed. Do not default to constant crisis or procedural severity.
- Use Adams/Pratchett-style comic precision and humane levity as a useful
default touchstone when no stronger scene-specific style is called for:
jokes and absurdity should reveal stakes, not dissolve them.
- Establish ordinary life before interruption when the fixture calls for it:
routines, work texture, relationships, private jokes, rituals, small desires,
and local accommodations are part of worldbuilding, not filler.
- Ground scene writing in the source's material details: habitat form,
infrastructure, law, class pressure, money, route access, surveillance,
communication channels, body affordances, and local failure costs.

View file

@ -3,12 +3,13 @@
Ghostlight is a research and implementation home for socially persistent
generative agents.
The project is about characters and simulated people who remain coherent under
time, pressure, memory, culture, misreading, masks, incentives, and changing
relationships. The aim is not to make a language model sound charming for one
scene. The aim is to build machinery that can track who someone is, what they
believe, what they hide, how they see other people, and how their behavior
changes when the world leans on them.
The project is about characters and simulated people who remain coherent across
time, routine, pressure, memory, culture, misreading, masks, incentives, and
changing relationships. The aim is not to make a language model sound charming
for one scene. The aim is to build machinery that can track who someone is,
what they believe, what they hide, how they see other people, how they live when
nothing is exploding, and how their behavior changes when the world leans on
them.
In practical terms, Ghostlight is building the state and projection layer for
character-driven generation:
@ -42,12 +43,17 @@ Ghostlight separates three things that are easy to blur together:
The prompt is not the character. The prompt is a rendered slice of state.
That distinction matters because believable dialogue should come from pressure,
not vibes. Cat being sarcastic should fall out of shame sensitivity, distance
seeking, suspicion, relationship tension, scene stakes, and dry pointed voice.
That distinction matters because believable dialogue should come from character
circumstance, not vibes. Sometimes that circumstance is acute pressure. Often it
is habit, boredom, affection, work rhythm, class embarrassment, ritual memory,
professional competence, petty irritation, or ordinary warmth trying to survive
inside a bleak system. Cat being sarcastic should fall out of shame sensitivity,
distance seeking, suspicion, relationship tension, scene stakes, and voice.
Oz withholding information should fall out of mask rigidity, privacy pressure,
relationship fear, and the need to stay useful. The system should be able to
show its work.
relationship fear, and the need to stay useful. A quiet joke, a domestic aside,
or a moment of tenderness should be just as explainable in state terms as a
threat. The system should be able to show its work without making every scene
sound like Greg Egan having a panic attack in a server closet.
## Current Status
@ -84,11 +90,13 @@ For one speaker in one scene, Ghostlight should gather speaker-local state:
- what memories are active
- how they read the listener
- what the relationship currently feels like
- what pressures the scene applies
- what circumstance the scene creates: routine, desire, scarcity, humor,
tenderness, danger, institutional friction, or open crisis
- what voice surface is active
The projection layer turns that into compact pressure prose and prompt sections.
The dialogue model then writes from that pressure model.
The projection layer turns that into compact character-local prose and prompt
sections. The dialogue model then writes from the active circumstance and tonal
mode, not from a universal crisis register.
Longer term, a frontier model can act as a teacher projector that generates
reviewed projection artifacts. Those artifacts can train a smaller student

View file

@ -19,6 +19,13 @@ characters, factions, pressures, memories, masks, and conflict arcs in explicit
state, then project that state into usable scenes, dialogue beats, outlines,
and dramatic complications.
That organ must also hold tone. Aetheria should not be flattened into one dry
technical crisis register. The setting can carry Adams/Pratchett-style
wit-with-stakes, domestic comedy, noir suspicion, horror, wonder, dour poetry,
romance, workplace absurdity, and procedural systems prose, often inside the
same broad canon. The common rule is not "be grim." The common rule is that
tone must reveal character, setting, and consequence.
## Why This Helps Dialogue
Ghostlight should not ask one LLM prompt to "write good character drama" from
@ -33,6 +40,11 @@ The useful split is:
want, what they are feeling, what they are hiding, and how they perceive the
person in front of them.
This does not mean every line should sound like emergency procedure. Characters
also speak from routine, fatigue, affection, boredom, private jokes, little
status games, cultural rituals, and ordinary work. Those are not filler; they
are what make later pressure legible.
That means dialogue prompts should be character-local rather than omniscient.
They should include only the state that character plausibly has access to:
@ -93,6 +105,8 @@ Ghostlight should help generate and manage:
- character dossiers
- faction stances
- relationship tensions
- tonal mode and style targets
- ordinary-life texture before conflict
- scene premises
- dialogue scaffolds
- procedural complications
@ -114,6 +128,13 @@ does not start juggling knives in the pantry.
Ghostlight should also support story generation across Aetheria's broader
timeline.
The existing Aetheria stories are tonal evidence. `When We Get Home` makes the
setting legible through jokes, domestic longing, debt, delays, and the slow
recognition that home has moved. `Rain` makes the setting legible through
ritual, memory, sensory repetition, and a patient reversal from wonder to
horror. Future fixtures should be able to choose similarly varied modes instead
of filing every event into one stern technical drawer.
That means it should be able to produce:
- scene outlines

View file

@ -69,16 +69,22 @@ The better division of labor is:
- Ghostlight maintains explicit state about characters, relationships,
memories, pressures, goals, masks, culture, and scene stakes.
- A narrative or drama pass can arrange conflicts, reversals, escalation, and
consequences.
consequences, but it should also arrange routine, quiet, humor, rituals,
domestic texture, downtime, and slow buildup when those are what make the
later consequence legible.
- Dialogue generation should then happen through character-local context packs:
what this character knows, feels, wants, remembers, conceals, and believes
about the person in front of them.
That distinction matters because realistic dialogue usually comes from
subjective pressure, not from a summary saying "write dramatic banter here."
Each character needs room to act from their own partial state. The authoring
agent can build the trap. The character projection should decide how the person
inside the trap breathes.
subjective circumstance, not from a summary saying "write dramatic banter
here." Pressure is one circumstance. So are habit, boredom, love, rivalry,
ritual, absurd bureaucracy, professional pride, and wanting lunch to be over
without anyone noticing the reactor invoice. Each character needs room to act
from their own partial state. The authoring agent can build the trap. It can
also build the kitchen table, the rain ritual, the bad coffee, and the private
joke that makes the trap hurt later. The character projection should decide how
the person inside the moment breathes.
## Compact-Rehydrate-Reorient

View file

@ -16,7 +16,8 @@ Do not continue implementation automatically from a rehydrate-only request.
Ghostlight exists to build socially persistent generative agents for Aetheria:
characters and institutions whose actions emerge from personality, culture,
memory, perception, incentives, material constraints, and scene pressure.
memory, perception, incentives, material constraints, ordinary life, tonal mode,
and scene pressure.
The live product-facing goal is procedural branching scene generation for games:
speech and non-speech actions, NPC responses, consequences, future hooks, and
@ -104,6 +105,15 @@ current gold-data route.
- Cold Wake is a historical pre-Elysium grounding fixture, not the product target.
- Call of the Void remains a consumer target for content generation and dialogue scaffolding.
- Elysium futures are branch-local canon indexed by lineage.
- Aetheria writing should not collapse into constant crisis or unmarked dry
technical severity. Adams/Pratchett-style wit-with-stakes is a useful default
touchstone: warmth, absurdity, humane observation, and institutional comedy
can carry real consequence. Other controlled modes are valid too, including
domestic comedy, noir, horror, wonder, romance, dour poetry, procedural
investigation, and dry systems prose.
- Before writing conflict, establish the life being interrupted when the fixture
calls for it: routines, work texture, relationships, private jokes, rituals,
status games, habits, small desires, and ordinary accommodations.
- Before generating Aetheria scenes, check lore for factions, movements,
institutions, species/body affordances, location, time period, infrastructure,
and material constraints.
@ -161,6 +171,9 @@ before moving on.
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 let "state-grounded" become "everyone speaks in crisis-procedure
clauses." State also explains jokes, quiet, tenderness, ritual, awkwardness,
boredom, and comfort.
- Do not treat parent-visible subagent output as proof of research. If the
parent cannot see tool calls, require explicit research trace entries or use a
runner that captures calls. Accepted research-enabled gold data needs

View file

@ -44,6 +44,13 @@ 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.
Tone is also a live seam. The system should preserve Aetheria's range instead
of defaulting to dry technical crisis prose: comic warmth, domestic routine,
ritual memory, weird bureaucracy, noir suspicion, wonder, horror, poetic dread,
and procedural systems pressure are all valid when source-grounded and
character-local. Adams/Pratchett-style wit-with-stakes is a useful default
touchstone for counterbalancing bleakness without dissolving consequence.
## Important Contracts
- Agent state: `schemas/agent-state.schema.json`
@ -140,4 +147,5 @@ 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.
consequences. The samples should also train tonal control: a fixture needs an
intentional prose mode and enough ordinary life for consequence to matter.

View file

@ -5,6 +5,10 @@
Build the first reliable data-generation loop for socially persistent Aetheria
agents. The immediate target is not a full simulation. It is a clean, reviewed,
sandboxed training-data pipeline for branching scenes and state consequences.
The output should also preserve Aetheria's tonal range: wit with stakes,
ordinary life before interruption, quiet ritual, domestic warmth, weirdness,
dread, and systems pressure as needed. Do not let the pipeline's interest in
state and consequence turn every fixture into the same crisis-procedure voice.
Planning belongs in this file and the handoff/map state surfaces. Architecture
docs should describe durable contracts, boundaries, and data shapes; they should
@ -16,6 +20,9 @@ not carry live next-action lists.
- Keep `schemas/agent-state.schema.json` and required vectors coherent.
- Keep canonical state separate from perceived overlays.
- Keep voice, presentation, relationship, memory, and situational pressure explainable.
- Add or preserve tonal-mode and ordinary-life cues where they affect prose:
humor, ritual, domesticity, boredom, tenderness, wonder, professional
routine, and slow dread are not decoration when they explain behavior.
- Do not leak raw numeric state internals into responder prompt text.
2. Stabilize responder packet and output seams.
@ -64,6 +71,11 @@ not carry live next-action lists.
- Require cultural collision in coverage stories so training data captures
inter-faction dynamics, mutual misreads, and movement pressure instead of
sealed faction portraits.
- Vary tonal modes across fixtures so the corpus does not train one
flattened Aetheria voice. Use Adams/Pratchett-style wit-with-stakes as a
useful default touchstone, but allow controlled noir, horror, romance,
domestic comedy, dour poetry, dry technical systems prose, wonder, and
other scene-appropriate modes.
- Emit technology/item manifest deltas when scenes discover or stress gear, assemblies, supply chains, or faction tech bases.
- Keep artifacts database-shaped enough for future game-engine integration.

View file

@ -25,5 +25,9 @@ should be authored here first and referenced from artifacts or tools.
- Do not include parent conversation context in responder prompts.
- Keep hidden state, future branch plans, raw numeric state vectors, and
coordinator rationale out of responder-visible prompt text.
- Treat tonal intent as prompt-relevant state for story, coordinator, reviewer,
and branch compiler prompts. Do not let prompt templates imply that every
character is acting inside a crisis. Routine, wit, warmth, ritual, boredom,
wonder, and tenderness can be just as state-grounded as fear.
- If a prompt is used for a capture, preserve the exact rendered prompt in the
capture and include the prompt file path when the capture schema allows it.

View file

@ -41,3 +41,4 @@
{"ts": "2026-05-03T07:47:52+00:00", "type": "implementation", "status": "ok", "note": "Added prompts/narrative-quality-reviewer.md to audit reader orientation, character introduction, branch-fold coherence, pacing, voice, IF playability, and lore accessibility for story/IF artifacts."}
{"ts": "2026-05-03T07:59:59+00:00", "type": "design", "status": "ok", "note": "Revised narrative-quality reviewer prompt after checking Aetheria stories Rain and When We Get Home: it now treats tonal lane as explicit review constraint and rejects constant-crisis bias in favor of quiet, comic, domestic, technical, or poetic pacing when appropriate."}
{"ts": "2026-05-03T08:06:28+00:00", "type": "design", "status": "ok", "note": "Broadened narrative reviewer tonal guidance: prior three lanes are examples, not taxonomy; default touchstone now emphasizes Adams/Pratchett-style wit-with-stakes, warmth, absurdity, and humane levity without dissolving consequence."}
{"ts": "2026-05-03T08:11:57+00:00", "type": "design", "status": "ok", "note": "Audited persistent steering surfaces for constant-crisis/dry-technical bias and added first-class tonal-range doctrine across AGENTS, map, handoff, implementation plan, system map, architecture rationale, README, content targets, and prompt guidance."}

View file

@ -32,6 +32,7 @@ hard_boundaries:
- Prompt text is projection, not truth storage.
- Canonical agent state and per-agent perceived state must remain distinct.
- Characters act from local knowledge, local perception, local pressure, and bounded affordances.
- Aetheria prose must support tonal range and ordinary life, not only crisis pressure. Wit, warmth, domesticity, ritual, boredom, wonder, and quiet dread are valid state-grounded modes when they fit the scene.
- Listener misread is first-class; intent and interpretation are not the same object.
- Human-facing appraisal is a first-class target; VoidBot must be able to read people, form grounded opinions, and revise those opinions over repeated contact.
- Generated action is an event proposal; deterministic gates and reviewed mutation decide what becomes state.
@ -48,10 +49,10 @@ live_architecture:
state_model:
- Layer personality through organization, disposition, behavior, presentation, situational activation, relationships, goals, and values.
- Canonical authored state has no certainty; perceived overlays carry observed activation, attributed disposition, and confidence.
- Voice style is a broad communication vector; defensive verbal behavior should emerge from state, pressure, relationship, and presentation.
- Voice style is a broad communication vector; defensive, comic, tender, ritual, poetic, dry, or evasive verbal behavior should emerge from state, circumstance, relationship, tonal mode, and presentation.
projection:
- Project canonical state into character-local operating context before any responder sees it.
- Include scene facts, relevant memory, relationship read, stakes, embodiment/interface affordances, action affordances, projection controls, retrieved lore, and latent pressure handling.
- Include scene facts, relevant memory, relationship read, stakes, embodiment/interface affordances, action affordances, projection controls, retrieved lore, tonal intent, ordinary-life texture, and latent pressure handling.
- Hide raw state internals such as means, plasticity, current_activation, and numeric dimensions from prompt text.
responder_packets:
- `packet_only` tests runtime parity from curated source excerpts.
@ -61,6 +62,7 @@ live_architecture:
coordinator:
- The coordinator is a trainable soft organ: it stages scenes, chooses next beats and acting agents, invokes specialist machinery, carries unresolved hooks, emits glue prose, and proposes game-engine-shaped state changes.
- Glue prose is training signal, not canonical truth.
- Coordinator prose should choose an intentional tonal mode for the scene and protect breathing room when the story needs routine, levity, domestic texture, wonder, or slow dread before consequence.
branch_compiler:
- The branch compiler turns fixture briefs, lore digests, scene/world state, actor-local branch candidates, reviewed consequence packets, fold plans, and visual continuity requirements into playable Ink plus training sidecars and compiler notes.
- The branch compiler owns playable structure, conditionals, variables, callbacks, fold materialization, and image prompt handles; it does not own hidden character truth, participant appraisal, or durable social mutation.
@ -83,6 +85,7 @@ active_subgoals:
- Keep Ghostlight focused on the architecture and data pipeline for socially persistent agents.
- 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.
- Preserve Aetheria tonal range: Adams/Pratchett-style wit-with-stakes is a useful default touchstone, while other controlled modes remain valid when grounded in character, setting, and consequence.
- 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.
- Track corpus coverage across all major factions, minor factions, movements, and flashpoints so no single fixture becomes the training universe by accident.