Move agent architecture docs into Ghostlight

This commit is contained in:
Meta Vens 2026-05-01 18:25:15 +01:00
commit 895723be30
13 changed files with 3136 additions and 26 deletions

View file

@ -31,6 +31,8 @@ beautiful little superstition.
live repo machinery
- `notes/ghostlight-implementation-plan.md`: current larger-organ sequence
- `notes/architecture-rationale.md`: why the state split exists at all
- `docs/research/`: research backlog and modeling notes
- `docs/architecture/`: world, state, and prompt-system architecture notes
- `tools/ghostlight_state.py`: compact state helper
- `tools/ghostlight_prepare_compaction.py`: pre-compaction audit helper
@ -45,3 +47,12 @@ npm run state:prepare-compaction
If you are steering the repo through Codex, read `AGENTS.md` first. That is the
operating manual. The README is just the sign on the door.
## Related Docs
- `docs/research/things-to-steal.md`
- `docs/research/personality-model-roadmap.md`
- `docs/research/personality-signal-classifier.md`
- `docs/architecture/agent-state-distributions-and-prompt-projection.md`
- `docs/architecture/aetheria-persistent-agent-architecture.md`
- `docs/architecture/aetheria-cold-wake-mvp-scenario.md`
- `docs/architecture/aetheria-agent-population-model.md`

View file

@ -0,0 +1,384 @@
# Aetheria Agent Population Model
This note explains how a socially dense Aetheria region can feel inhabited
without simulating every implied person as a fully persistent agent.
The short answer is that population and agency are not the same thing.
A corridor can imply hundreds of thousands or millions of people while only a
few dozen individuals need to exist as full long-lived social nodes.
## Design Goal
We want a region that feels:
- crowded
- economically alive
- politically pressured
- socially stratified
- full of rumor, labor, traffic, and institutional drag
We do **not** want:
- ten thousand expensive little chatbots idling in a warehouse
- fake scale achieved by silent cardboard extras
- a world where only named characters seem to exist
So the population model needs to separate:
- who exists as a persistent social mind
- who exists as a grouped social force
- who exists only as local scene color
- how offscreen population pressure becomes visible
## Four Population Layers
The MVP should use four layers.
### 1. Persistent Agents
These are the full agents.
They have:
- identity
- memory
- relationships
- goals
- obligations
- changing stance toward institutions and people
These are the characters who accumulate history and can plausibly remember what
happened last week, who lied to them, who saved them, and which institution
just humiliated them.
Examples in the Cold Wake slice:
- a navigator council delegate
- a Lightsail dispatcher
- a dock labor organizer
- an insurer adjudicator
- an Aya sanctuary coordinator
- a pirate broker
- a cutout cognition vendor
- a Cryonix or Rossum technical rep
These are not the population. They are the pressure points where
population-scale systems become legible.
### 2. Cohorts
Cohorts are grouped social bodies that matter but do not need full individual
cognition.
They should represent classes of people whose collective condition can change
over time.
Examples:
- dock labor shifts
- convoy deck crews
- insurer back-office clerks
- sanctuary residents
- clinic queues
- pirate route crews
- private security contractors
- debt pilgrims
- repair-bay technicians
- refugee transit waves
Each cohort should have state such as:
- size
- location
- stress
- trust in institutions
- rumor exposure
- faction sympathy
- fatigue
- scarcity pressure
- protest potential
- desertion / mutiny risk
This is how the world gets to have labor, crowds, queues, shortages, and
unrest without pretending every forklift operator needs a hundred memories and
a private speech cadence.
### 3. Ephemeral Scene Actors
These are temporary people instantiated only when a scene needs bodies,
witnesses, clerks, mechanics, medics, bystanders, or small bits of friction.
They should have:
- a role
- a scene goal
- local awareness of current events
- a little voice
- minimal short-term memory
They should **not** be treated as deep persistent minds by default.
Examples:
- a berth marshal waving traffic through a panic checkpoint
- a frightened passenger arguing with a clerk
- a dockhand spreading a fresh rumor
- a medic triaging one more body than the shift can really handle
- a sanctimonious insurer junior who disappears once the scene ends
If an ephemeral keeps mattering, the system can promote them into a persistent
agent later.
### 4. Ambient Population Pressure
This is the layer that makes the world feel bigger than the visible cast.
It is not represented by people at all. It is represented by signals.
Examples:
- berth queues growing or shrinking
- cargo backlogs
- corridor delay maps
- sanctuary occupancy
- clinic wait times
- food prices
- insurance repricing
- labor slowdown notices
- protest bulletins
- casualty lists
- rescue requests
- media chatter
- tribunal rulings
- route closures
This is where the implied millions live in practice.
## How The Layers Interact
The model should feel like this:
- ambient pressure changes cohort state
- cohort state changes the incentives around persistent agents
- persistent agents make decisions that alter ambient pressure
- ephemeral actors express the local scene consequences
That loop lets a single adjudicator decision or pirate rescue cascade outward
into visible population-level effects without simulating every single person
affected.
## Why A Few Dozen Persistent Agents Is Enough
Because the simulation is not trying to create a census. It is trying to create
a socially legible machine.
In a corridor like Ganymede, one persistent agent can sit at the hinge point
for a great many people:
- one dispatcher affects dozens of ships and hundreds of crew
- one sanctuary coordinator gates a network under pressure from thousands of vulnerable lives
- one insurer official can reprice route access across the corridor
- one navigator delegate can legitimize or delegitimize rescue, inspection, and route closure
- one pirate broker can redirect shadow traffic across many dependent cells
The persistent cast is therefore best understood as:
- decision-makers
- bottlenecks
- brokers
- organizers
- interpreters of institutional rules
- people with unusual leverage over larger flows
They are the knots in the rope, not every fiber.
## Cohort Model
Cohorts should be tractable and opinionated.
Each cohort record should include:
- `cohortId`
- `type`
- `sizeBand`
- `region`
- `factionAffinity`
- `currentPressures`
- `resourceDependency`
- `confidence`
- `grievance`
- `discipline`
- `mobilizationPotential`
- `recentIncidents`
Useful cohort types for Cold Wake:
- dock labor
- convoy crews
- sanctuary residents
- insurer workforce
- pirate crews
- repair-yard workers
- shadow clinic clients
- private security detachments
- local merchant bodies
- transient passenger pools
Cohorts do not need rich dialogue. They need believable collective reactions.
## Ephemeral Actor Rules
Ephemeral actors should be generated from:
- scene location
- cohort state
- current world pressures
- factional tone
- time of day / shift
That means the same dock concourse can produce different incidental people
depending on whether the corridor is calm, panicking, starving, over-inspected,
or politically inflamed.
Ephemerals should support:
- local exposition
- scene texture
- small moral friction
- witness testimony
- short disputes
- visible consequences
They should expire cheaply unless the world keeps pulling them back into
relevance.
## Promotion Rules
Some ephemerals should be promoted into persistent agents when:
- a player or major agent forms an ongoing relationship with them
- they become central to repeated incidents
- they accumulate leverage or notoriety
- they are the surviving witness to something politically explosive
- they become the face of a cohort-level shift
Promotion is how the world grows specific people organically instead of
requiring the whole cast to be authored up front.
## Population Signals That Make A Place Feel Busy
If the corridor is meant to feel inhabited, the system should constantly
surface signs of offscreen life.
That means scenes, dashboards, and local UI should be fed with things like:
- arrivals and departures
- route advisories
- late manifests
- dockside rumor digests
- occupancy warnings
- strike whispers
- inspection delays
- tribunal alerts
- debt-claim notices
- casualty and rescue numbers
- commodity shocks
- public feeds and scandal bursts
The player does not need to meet everyone. They need to feel surrounded by
consequences.
## Cold Wake-Specific Population Recipe
For the first slice, a sensible starting mix is:
- **20-40 persistent agents**
- **10-25 cohorts**
- **0-20 ephemerals per active scene**, generated on demand
Persistent clusters:
- Navigators
- Lightsail
- PSC / insurers
- Aya
- pirates
- cutout vendors
- local technical reps
Cohorts:
- dock labor
- convoy crews
- sanctuary overflow
- clinic queues
- inspector pool
- security contractors
- pirate cells
- repair techs
- cargo handlers
- passengers / transients
That is enough to make the world feel socially populated without turning the
runtime into a screaming furnace.
## Simulation Tier Mapping
The population layers should map onto the simulation tiers cleanly.
### Tier 0
- dormant persistent agents
- slowly changing cohorts
- coarse ambient indicators
### Tier 1
- active persistent agents making strategic moves
- cohorts shifting under pressure
- route, labor, and sanctuary state updating
### Tier 2
- a handful of persistent agents in scene
- relevant ephemerals generated for the local moment
- cohort state informing mood, risk, and crowd behavior
### Tier 3
- full interactive dialogue for the most relevant persistent agents
- ephemerals only as needed for texture or interruption
- world pressure visible through feeds, queues, alerts, and consequences
This keeps expensive cognition focused where it matters.
## What Not To Do
Do not:
- simulate every resident as a full agent
- let cohorts become inert statistical wallpaper
- use ephemerals as consequence-free decoration
- make named agents behave as if the rest of the world does not exist
- make ambient scale purely visual with no mechanical pressure
That path leads to a world that looks crowded and feels dead.
## Success Criteria
The model is working if:
- the player can feel institutional pressure without being told about it in lore prose
- small decisions by key agents visibly affect larger populations
- cohorts can become restless, relieved, fearful, or politically active
- scenes feel crowded when they should
- the same corridor can feel different under calm, panic, shortage, scandal, or repression
- the named cast feels embedded in a larger society rather than posed in a diorama
The real test is simple:
If the player leaves the room, do they still believe thousands of people are
having a bad day because of what just happened?
If yes, the world is inhabited.

View file

@ -0,0 +1,451 @@
# Aetheria Cold Wake MVP Scenario
This is the first buildable scenario brief for the persistent-world agent work.
The broader machine lives in
[Aetheria Persistent Agent Architecture](./aetheria-persistent-agent-architecture.md).
The population scaling answer lives in
[Aetheria Agent Population Model](./aetheria-agent-population-model.md). This
document is the narrower, dirtier thing: one bounded slice of the setting that
is small enough to simulate and rich enough to misbehave.
## Premise
The MVP should center on **Ganymede orbital docks and one outward Jovian convoy
corridor during the Cold Wake Panic**.
That gives us:
- a physically bounded region
- a politically meaningful choke point
- factions with incompatible incentives
- local social density
- enough offstage Sol to make every local event matter
The corridor is not just transit. In Aetheria, it is law, trust, debt, rescue,
smuggling, insurance, and quiet warfare wearing the same pressure suit.
## Why This Slice
This slice works because the lore already piles several important systems on top
of one another:
- **Ganymede** sits at the overlap of route law, dock labor, shipbuilding support, and outer-system logistics.
- the **Ganymede Route Compact** is a foundational moment in the rise of the Cetacean Navigators and in corridor governance itself
- **Jovian space** already contains both high-legibility convoy culture and murkier gray-market cognition and pirate traffic
- **Cold Wake Panic** turns thermal stealth, quiet-running, and vessel classification into a live political crisis
That means we do not have to invent tension from scratch like desperate little
frauds. The world already wants to fight.
## Flashpoint
The live scenario should be built around **Cold Wake Panic** rather than the
older Ganymede Route Compact.
The Compact is excellent historical texture and local civic mythology. It tells
us what people in this region think corridor legitimacy is supposed to mean.
Cold Wake Panic is the better active crisis.
During the panic:
- insurers realize thermal posture has become a strategic problem
- permit categories no longer describe the real battlespace
- quiet-running ships can stay cold long enough to make first detection politically explosive
- Rossum & Douglas and Cryonix become important to certified quiet-running envelopes and heat-debt tolerances
- privateers, security buyers, and corridor authorities all start circling each other
That is ideal proof-of-concept material because it creates immediate uncertainty
with economic and moral consequences.
## Bounded World Slice
The MVP world should include four linked spaces:
1. **Ganymede Orbital Docks**
- main social hub
- cargo handling
- navigator councils
- insurer desks
- PSC corridor administration
- labor syndicate politics
2. **Convoy Corridor Segment**
- one actively monitored route
- route hazards
- rescue obligations
- stealth / heat-debt / interdiction pressure
3. **Callisto Spillover Zone**
- shadow logistics
- cutout research
- pirate brokerage
- gray-market cognition work
4. **Relay Dark / Tribunal Layer**
- distributed insurer rulings
- permits
- sanctions
- misclassification disputes
- offscreen political leverage made locally visible
This gives us one city-like social knot, one moving logistics lane, one
shadow-edge, and one bureaucratic abstraction layer that can still hurt people.
## Historical Texture
The old wound under the scenario is the **Ganymede Route Compact**.
Locally, people should treat it the way cities treat founding strikes,
rebellions, or constitutional bargains:
- navigators remember it as the point where corridor labor stopped being property and became politics
- Lightsail and dock operators remember it as the point where reliability got institutional teeth
- insurers remember it as the moment rescue and hazard governance became unavoidable
- hard-sovereignty actors remember it as the moment routes stopped behaving like obedient assets
The scenario should not reenact the Compact. It should let people invoke it.
That gives local conversation and factional self-justification a spine.
## Local Factions
The MVP does not need every solar faction embodied at full fidelity, but they
should all still exist as economic and political pressure.
### Fully Embodied In The Slice
- **Cetacean Navigators**
- route stewardship
- rescue legitimacy
- corridor governance
- **Lightsail Express**
- fleet operations
- convoy reliability
- practical logistics power
- **Pan-Solar Consortium / PSCC / insurer apparatus**
- corridor legality
- clearing
- sanctions
- classification and risk pricing
- **Aya Collective**
- sanctuary intake
- clinics
- food and repair capacity
- moral pressure against procedural cruelty
- **Pirate Coalition**
- shadow circulation
- escape routes
- black cargo
- opportunism, rescue, and predation in unstable proportions
### Present Through Cutouts, Contracts, Or Local Representatives
- **Cognitum**
- gray-market cognition work
- routing firmware
- clinic/bounty harvesting
- deniable strategic minds
- **Rossum & Douglas**
- quiet-running certification
- thermal-envelope expertise
- **Cryonix**
- signature-managed skins
- heat-debt infrastructure
- **AstroDyne**
- rough frontier repair culture
- cargo and salvage spillover
- **GeneSys**
- old uplift lineage hanging over Navigator politics
### Offstage But Causally Present
- Sol Dominion
- Zhestokost
- Lucent Media
- Framgang
- Baseline League
- other major powers and movements
These do not need local populations in the MVP. They need to arrive as policy,
money, media shocks, security contracts, doctrine, and cargo priority.
## Why Cognitum Fits
Cognitum is not explicitly written as the overt local ruler of Ganymede. Good.
That would be a little too neat and a lot less rancid.
What the lore gives us instead is better:
- Ganymede and Callisto host shadow economies, cutout research, and gray-market cognition work tied back to the Cognitum and pirate logistics
- Cognitum specializes in deniable cognition products, routing tools, and firmware derived from uploaded minds
- it hunts for minds with traits including navigational instinct
- after public scandal it hardens its cutout structure and pushes more work through shell firms, clinics, and nominally separate vendors
So the local Cognitum footprint should look like:
- bounty clinics
- shell routing consultants
- firmware brokers
- black-budget procurement intermediaries
- suspiciously capable quiet-running optimization suites
- trafficked or coerced cognition hidden inside legal cargo categories
That means the players can discover Cognitum through traces, testimony,
invoices, wreckage, and people breaking under pressure instead of through a
villain sign hanging over a lobby.
## Core Tensions
The slice works if every faction is defending something real and mutually
incompatible.
### Navigators
Need routes to remain governable through trust, rescue obligation, and
stewardship.
They fear:
- routes becoming pure battlespace
- security actors replacing reciprocity with force
- cognition exploitation hiding inside logistics systems
### Lightsail
Needs ships to keep moving and reputation to remain intact.
They fear:
- insurer panic
- corridor shutdowns
- being forced to choose between lawful compliance and actual reliability
### PSC / insurers
Need categories, permits, and thermal classifications to mean something again.
They fear:
- political scandal caused by misclassification
- uncontrollable stealth escalation
- loss of leverage if corridor law looks absurd
### Aya
Needs sanctuary and care systems to survive without being flooded past
capacity.
They fear:
- rescue obligations collapsing into selective abandonment
- every crisis being solved by securitization
- becoming a moral fig leaf for systems they cannot actually stop
### Pirates
Need borders to remain permeable and hard closures to fail.
They fear:
- route legibility becoming total
- sanction and thermal law eliminating the shadow economy
- rival pirate or privateer factions destroying the social trust they still need to operate
### Cognitum And Friends
Need deniable cognition work to remain profitable and plausibly untraceable.
They fear:
- personhood claims attaching to firmware
- exposure of coercive routing or quiet-running products
- corridor governance deciding some kinds of optimization are too morally radioactive to permit
## Starting Crisis
The scenario should open with a single destabilizing incident:
- a quiet-running vessel appears in or near the corridor under classifications that do not line up
- thermal telemetry is contradictory
- insurer desks freeze
- some local actors think the ship is pirate
- some think it is a security black project
- some think it is a refugee or sanctuary run
- some think its routing behavior implies firmware or overclocked cognition that should not legally exist
The point is not to start with the answer. The point is to start with a problem
that every faction is forced to interpret through its own incentives.
## Persistent Agent Cast
The MVP does not need thousands of agents. It needs a small cast with enough
institutional reach and private mess to create chain reactions.
Target range:
- **20-40 persistent agents**
Suggested clusters:
- navigator council delegates
- convoy captains
- Lightsail dispatchers
- dock labor organizers
- insurer adjudicators
- PSC sanctions and permit clerks
- Aya sanctuary coordinators
- Aya medics / rights advocates
- pirate brokers
- pirate rescue crews
- repair-yard operators
- shell-company cognition vendors
- one or two Rossum / Cryonix technical reps
- investigators, rumor traders, or media intermediaries
Every one of these agents should have:
- obligations
- debts
- loyalties
- private fears
- a view of the Compact myth
- an opinion on whether rescue overrides classification
## Resources And State Variables
The simulation should track a few things explicitly instead of trying to
simulate the whole economy like an idiot.
### Material
- cargo throughput
- fuel / heat debt
- med capacity
- sanctuary bed capacity
- repair queue
- available escort / security coverage
### Social / Political
- corridor trust
- insurer confidence
- sanction pressure
- pirate route permeability
- public scandal temperature
- evidence of cognition abuse
- legitimacy of navigator stewardship
### Agent-Level
- trust
- resentment
- obligation
- leverage
- fear
- secrecy burden
These are enough to make choices matter without drowning the system in
decorative bookkeeping.
## Event Escalation Ladder
The world should have a simple ladder that can climb if humans and agents keep
making bad decisions.
1. **Misclassification**
- one vessel cannot be cleanly categorized
2. **Risk Repricing**
- insurers tighten rules and drive panic behavior
3. **Route Friction**
- delays, refusals, inspections, and rescue disputes spread
4. **Exposure**
- evidence of gray-market cognition or cutout procurement emerges
5. **Faction Split**
- Navigators, PSC, Aya, and Lightsail stop agreeing on minimum procedure
6. **Security Turn**
- interdiction, seizures, or retaliatory piracy escalate
7. **Legitimacy Crisis**
- the corridor stops believing in its own rules
The simulation does not need a single golden path. It needs enough pressure
that different policies lead to different failures.
## Human Participation
Humans should not be the sole drivers of history here. They should be able to
disturb an already-running machine.
Possible player roles:
- dock worker
- independent courier
- sanctuary volunteer
- pirate-adjacent fixer
- insurer clerk
- navigator aide
- repair tech
- contract investigator
The important thing is that players should affect:
- who gets trusted
- what evidence circulates
- who gets rescued
- whether secrets stay secret
- which institutions lose face
They should not be treated like prophesied little suns for whom society
politely pauses.
## MVP Build Order
If we build this, the order should be:
1. **Static scenario bible**
- locations
- factions
- route rules
- historical context
2. **Persistent state model**
- agents
- relationships
- route and resource variables
3. **Strategic tick simulation**
- movement
- cargo
- permits
- rescue and inspection pressure
4. **Scene simulation**
- disputes
- negotiations
- rumors
- sanctuary triage
5. **Live human interaction**
- dialogue
- persuasion
- betrayal
- evidence exchange
6. **Escalation events**
- panic waves
- seizures
- leaks
- route failures
The point is to make the corridor feel alive before trying to make it huge.
## Why This Is A Good First Swing
This slice is large enough to feel like civilization and small enough to reason
about.
It lets us prove:
- persistent agents with memories and goals
- factional friction that comes from incentives instead of plot rails
- human participation without player centrality
- offstage macro-factions influencing a local world
- morally ugly technology showing up as lived social pressure instead of wiki trivia
Most importantly, it gives us a world where the interesting question is not
"what quest does this NPC have for me?" but "what kind of order survives when
trust, law, care, and profit all want different things from the same
corridor?"

View file

@ -0,0 +1,574 @@
# Aetheria Persistent Agent Architecture
This is the companion document to
[Things To Steal](../research/things-to-steal.md).
For the first concrete slice, see
[Aetheria Cold Wake MVP Scenario](./aetheria-cold-wake-mvp-scenario.md).
For the scaling model that explains how a crowded region can be simulated with
a relatively small persistent cast, see
[Aetheria Agent Population Model](./aetheria-agent-population-model.md).
For the deeper note on how culture, personality distributions, memory, and
perceived state should project into prompts, see
[Agent State Distributions And Prompt Projection](./agent-state-distributions-and-prompt-projection.md).
That document is the theft list. This one is the machine we are trying to build
out of the stolen parts.
The target is a persistent world full of in-universe agents who:
- remember what happened
- want different things
- collide with each other because their values and interests do not line up
- interact with humans without collapsing into quest dispensers or therapy plushies
The important bit is that the world should not exist solely to flatter the
player. Humans are participants, not the sun around which every synthetic
little bastard must orbit.
## Design Goals
We want a world where:
- agents persist across sessions
- relationships evolve over time
- off-screen life continues
- conflict emerges from incentives
- the same character remains recognizably themselves under pressure
- human intervention can redirect events without erasing everyone else's agency
We do **not** want:
- a thin layer of LLM chat stapled onto static NPC behavior
- every character becoming a slightly differently dressed assistant
- a world that only animates when a human looks at it
- fake drama generated by random hostility instead of meaningful friction
## World Model
The world should be represented as a persistent simulation with five main state
layers:
1. **Canonical world state**
- locations
- factions
- resources
- institutions
- active conflicts
- world-level events
2. **Agent identity state**
- role
- background
- loyalties
- values
- aversions
- appetites
- speech style
3. **Agent dynamic state**
- current location
- current obligations
- current mood / pressure / fatigue
- current social exposure
- short-horizon plans
4. **Relationship state**
- trust
- respect
- resentment
- fascination
- fear
- obligation
- debt
5. **Memory state**
- episodic incidents
- scene summaries
- semantic profile conclusions
- relationship summaries
- active inferred goals and concerns
## Agent Stack
Each agent should be built from explicit layers instead of one giant persona
prompt.
### 1. Identity Layer
Relatively stable facts:
- who they are
- what they believe
- what they want to be seen as
- what they cannot tolerate
- what they are drawn toward
This should be treated as durable but revisable. Identity is not immutable, but
neither should it drift because the model got bored.
### 2. Relationship Layer
Every important other should have a relationship record, including:
- another agent
- a human player
- a group or faction
The record should track:
- trust
- respect
- resentment
- dependence
- attraction or fascination where relevant
- leverage
- recent unresolved incidents
The point is not to create one friendship slider like a life sim from 2004. The
point is to track the actual shape of a social bond.
### 3. Goal Layer
Agents need active goals, not just a biography.
Goals should be:
- private or public
- immediate or long-range
- strategic or emotional
- revisable in response to new information
A goal object should include:
- description
- priority
- time horizon
- emotional stake
- blockers
- dependencies
- known rivals
- likely next moves
### 4. Memory Layer
Memory should be split into:
- **working memory**
- current scene
- immediate context
- active inferences
- **episodic memory**
- timestamped incidents
- where, who, what, consequence
- **scene memory**
- distilled summaries of interactions
- **semantic memory**
- generalized beliefs about self, others, and world
- **relationship memory**
- remembered incidents and inferred patterns tied to a specific other
This is where the current interaction-memory work is directly useful.
### 5. Theory-of-Mind Layer
Each active interaction should have a lightweight scratchpad:
- what the other party likely wants
- what they fear
- what they think is happening
- what they are trying to get from me
- how confident we are about those guesses
This should be ephemeral and revisable. First impressions are useful, but
worshipping them is how systems become smug and stupid.
## Simulation Tiers
A persistent world cannot run every agent at full dialogue fidelity all the
time unless we want to immolate the hardware.
So the world should run on tiers.
### Tier 0: Dormant Persistence
For agents far away from any relevant action:
- maintain location, obligations, and long-horizon goals
- advance the clock coarsely
- update minimal state
### Tier 1: Strategic Tick
For agents in active regions but not in a live scene:
- update goals
- choose broad actions
- move, trade, coordinate, prepare, avoid, investigate
- write compact event records
### Tier 2: Scene Simulation
For agents inside a socially or narratively relevant scene:
- model who is present
- what each actor knows
- what each actor wants
- who is hiding what
- who is under pressure
- resolve short exchanges, negotiations, or conflicts
### Tier 3: Live Interactive Presence
For agents currently engaged with a human or central dramatic scene:
- full dialogue generation
- richer perception and recall
- fine-grained ToM updates
- explicit action selection tied to the live interaction
This tiered model is how we get a world that feels alive without paying for a
small national grid.
## Core Loop
At a high level, each active agent loop looks like:
1. Perceive
- scene state
- nearby actors
- new events
- human input
2. Interpret
- what matters
- what changed
- what this implies for goals and relationships
3. Update internal state
- working memory
- ToM scratchpad
- emotional pressure / stance
4. Retrieve memory
- relevant episodes
- relationship incidents
- semantic conclusions
5. Evaluate goals
- what is threatened
- what is newly possible
- what should take precedence now
6. Choose action
- speech
- movement
- trade
- alliance
- concealment
- refusal
- violence
- delay
7. Commit result
- world state update
- event log
- memory writes
8. Reflect later
- consolidate incidents
- revise relationship state
- revise semantic beliefs
- revise goals
## Relationship Engine
The relationship system should be treated as first-class infrastructure, not
decorative metadata.
Each relationship record should have:
- scores or categorical state for trust, respect, resentment, obligation, fear, fascination
- a short summary of the current relationship
- the unresolved incidents driving that summary
- known pressure points
- current access level
Example consequences:
- low trust + high obligation = brittle alliance
- low respect + high dependence = contemptuous reliance
- high fascination + high fear = unstable orbit
- high resentment + high access = sabotage risk
That is where drama comes from.
## Goal Engine
Goals should be generated and updated from:
- identity
- current needs
- memory
- faction context
- scene opportunities
- rival actions
We should distinguish between:
- **maintenance goals**
- survive
- maintain station
- avoid exposure
- **strategic goals**
- secure influence
- gain a resource
- undermine a rival
- **emotional goals**
- seek vindication
- restore dignity
- protect someone
- punish betrayal
Emotional goals matter because otherwise every character becomes a boring little
optimization engine with no pulse.
## Scene Engine
A scene should be an explicit object, not just "some chat happened."
A scene record should contain:
- location
- participants
- public stakes
- hidden stakes
- triggering event
- current tension axes
- known facts by actor
- unknown facts by actor
- scene outcome
Useful tension axes:
- status
- safety
- secrecy
- leverage
- affection
- ideology
- scarcity
The scene engine should decide:
- whether to escalate
- whether to de-escalate
- whether to fork into violence, bargain, confession, retreat, or deception
## Human Participation
Humans should be one force in the world, not the privileged audience every
character was born to adore.
That means:
- agents should remember human actions
- agents should form durable impressions of specific humans
- agents should change how much they reveal, trust, or tolerate
- agents should be willing to refuse or redirect
- agents should sometimes try to use the human for their own ends
Humans should be able to:
- build alliances
- ruin them
- become symbols
- become liabilities
- become prey
- become confidants
That is much more interesting than "click NPC, receive lore paragraph."
## Narrative Layer
We should not hard-script drama, but we should summarize it.
The narrative layer should:
- generate personal chronicles from actual events
- generate regional or faction news from high-impact changes
- generate rumor propagation when information is partial or distorted
The crucial rule:
- narrative is downstream of simulation
- not a replacement for simulation
## How The Current Supporting Stack Maps To This
The nice part is that some useful scaffolding already exists elsewhere.
Current useful pieces:
- Postgres-backed persistent state
- Qdrant-backed semantic retrieval
- interaction memory with behavioral dimensions
- source and lore indexing
- provider-side tool use
- owner-safe MCP patterns
What needs to be added:
- multi-agent world state store
- relationship graph
- explicit goal objects
- scene scheduler
- reflection/consolidation jobs for cast state
- tiered simulation runtime
- world-facing tools for observation and action
In other words, the existing conversational stack is already halfway to being a
mind-and-memory service. What it lacks is the world model and the cast
scheduler.
## Suggested Data Model
At minimum, the persistent store should grow tables or equivalents for:
- `world_regions`
- `world_events`
- `agents`
- `agent_identity_snapshots`
- `agent_dynamic_state`
- `agent_goals`
- `agent_relationships`
- `agent_memories_episodic`
- `agent_memories_scene`
- `agent_memories_semantic`
- `scene_instances`
- `scene_participants`
- `human_agent_interactions`
- `rumors`
Qdrant or equivalent can remain the retrieval layer for:
- episodic recall
- lore retrieval
- source grounding
- scene similarity
But the canonical social and world state should stay relational and
inspectable.
## Suggested Runtime Services
The world runtime should likely split into:
- **world scheduler**
- advances ticks
- activates scenes
- downgrades inactive agents
- **agent executor**
- runs an agent turn
- retrieves memory
- updates goals and relationships
- **scene resolver**
- handles active multi-party interactions
- **reflection worker**
- consolidates events into longer-term memory
- **narrative worker**
- writes chronicles and reports
That keeps the system legible instead of letting one monstrous worker do all
the thinking and all the plumbing at once.
## Failure Modes To Watch For
### 1. Everyone Becomes The Same Person
The agent voices flatten and all conflict becomes cosmetic.
Countermeasure:
- strong identity layer
- explicit value hierarchy
- personality consistency evaluation
### 2. Memory Becomes Sludge
Everything is stored, nothing is distinguished, retrieval turns to soup.
Countermeasure:
- tiered memory
- write-path filtering
- periodic consolidation
- clear forgetting policy
### 3. Drama Becomes Random Meanness
Characters feel chaotic rather than motivated.
Countermeasure:
- tie conflict to values, history, incentives, and pressure
### 4. Humans Break The World By Existing
Every character over-orients around the nearest player.
Countermeasure:
- keep off-screen goals live
- preserve non-human social gravity
- let humans matter without becoming the axis mundi
### 5. Sycophancy Wears A Costume
The bot sounds spicier but still validates whatever frame the user wants.
Countermeasure:
- explicit anti-sycophancy evaluation
- persistent relationship memory
- boundary-aware response guidance
## Build Order
### Phase 1: Cast Spine
- explicit agent identity dossiers
- explicit goal objects
- explicit relationship objects
- episodic + semantic memory split
### Phase 2: Social Simulation
- scene objects
- scene scheduler
- relationship updates from scenes
- goal revision from scenes
### Phase 3: Persistent World
- region state
- off-screen strategic ticking
- rumor/news propagation
- faction and economy hooks
### Phase 4: Human Integration
- live player-facing interactions
- persistent human reputation
- access, trust, and consequence systems
### Phase 5: Rich Narrative Surface
- chronicles
- regional reports
- replay and spectator tooling
## The Real Ambition
The ambition here is not "better NPC dialogue."
It is:
- a world with inhabitants
- inhabitants with memory
- memory that changes what they want
- wants that collide
- collisions that produce story
That is the actual prize. Everything else is just wiring and smoke.

View file

@ -0,0 +1,399 @@
# Agent State Distributions And Prompt Projection
This note describes how personality, culture, memory, and social context should
be represented in structured state, and how that state should be projected into
prompts for dialogue generation.
The central rule is simple:
- the prompt is not the source of truth
- the prompt is a temporary projection of deeper state for one turn or one scene
If we try to store the whole person in prompt prose, the system turns back into
theatrical mush. The real machine has to live in explicit state.
For the related notes on personality structure and dialogue classification, see:
- [Personality Model Roadmap](../research/personality-model-roadmap.md)
- [Personality Signal Classifier](../research/personality-signal-classifier.md)
## Core Principle
We should model agents through:
1. canonical latent state
2. perceived state for each agent
3. prompt projection from perceived state into live generation
That gives us:
- stable world truth
- subjective misunderstanding
- efficient live prompting without pretending the prompt is the database
## State Families
Each agent should have explicit state families, not one giant personality blob.
### 1. Underlying Organization
These are deep structural variables:
- self-coherence
- contingent worth
- shame sensitivity
- reciprocity capacity
- mentalization quality
- authenticity tolerance
- mask rigidity
- external-regulation dependence
These are slow-changing and often heavily shaped by early development and
repeated relational history.
### 2. Stable Dispositions
These are relatively persistent tendencies:
- novelty-seeking
- conformity
- status hunger
- risk tolerance
- sociability
- baseline threat sensitivity
- aesthetic appetite
- ideological rigidity
### 3. Behavioral Dimensions
These are first-class, not optional garnish. They belong in the rebuild from
the start:
- warmth
- drive
- grandiosity
- validation-seeking
- anxiety
- control pressure
- hostility
- suspicion
- rigidity
- withdrawal
- volatility
- attachment-seeking
- distance-seeking
These are the main handles for social behavior, institutional pressure, and
consumer or factional drift.
### 4. Presentation Strategy
These describe how the self is performed:
- charm
- compliance
- superiority
- detachment
- seductiveness
- competence theater
- moral theater
- strategic opacity
- cultivated harmlessness
These are not just decorative style. In some agents they are survival
machinery.
### 5. Situational State
These are transient activations:
- exhaustion
- scarcity pressure
- humiliation
- panic
- triumph
- grief
- overstimulation
- grievance activation
- acute shame
- perceived status threat
### 6. Relationship State
Per target person, faction, group, product domain, or institution:
- trust
- resentment
- dependence
- fear
- fascination
- obligation
- envy
- moral disgust
- perceived status gap
- expectation of care
- expectation of betrayal
### 7. Goals And Values
- active goals
- protected values
- emotional stakes
- tradeoffs
- unforgivables
## How To Represent Variables
Do not represent most of these as naked scalars if we want culture and history
to shape them properly.
A practical representation is:
- `mean`
- `certainty` or `concentration`
- `plasticity`
- `current_activation`
So instead of:
- `status_hunger = 0.72`
use something like:
- `status_hunger.mean = 0.72`
- `status_hunger.certainty = 0.81`
- `status_hunger.plasticity = 0.22`
- `status_hunger.current_activation = 0.89`
This gives us:
- stable baseline personality
- uncertainty and variation
- different rates of change
- scene-sensitive activation without rewriting the whole person every turn
## Culture As A Prior Field
Culture should not create clones. Culture should shape distributions.
For each culture, faction, class stratum, or institution, define:
- prior means over variables
- covariance structure
- reward gradients
- penalty gradients
- preferred masks
- taboo map
- prestige map
- default relationship posture to insiders, outsiders, rivals, inferiors, and superiors
- favored narrative scripts
This means a culture is not "people here are high in trait X."
It is more like:
- this culture rewards controlled presentation
- penalizes spontaneous vulnerability
- elevates status anxiety
- normalizes suspicion of low-status outsiders
- attaches prestige to procedural fluency and emotional composure
That is how a Dominion or Lucent style environment should work: as a field that
shapes what people become profitable or safe enough to be.
## Personal History As A Transform Layer
Culture is only the starting pressure.
Then the prior is transformed by:
- attachment history
- class position
- faction indoctrination
- educational shaping
- labor regime
- remembered humiliations
- remembered betrayals
- remembered rescue, admiration, abandonment, and obligation
So agent state should be read as:
- cultural prior
- transformed by developmental history
- transformed by remembered episodes
- currently activated by scene pressure
## Cohorts And Populations
The same representation can be lifted from individuals to populations.
A colony, district, labor bloc, or consumer cohort can have:
- shared priors
- institutional conditioning
- repeated collective incidents
- relationship state toward factions, employers, media ecosystems, and product classes
This is much better than Big Five-style population sliders because it lets us
model things like:
- prestige hunger
- fear-state compliance
- grievance accumulation
- attachment to institutions
- validation loops induced by media
- suspicion induced by audit or betrayal
That is much closer to Aetheria than a broad trait scaffold alone.
## Canonical State Vs Perceived State
The world should not only track what is true. It should track what each agent
thinks is true.
So we need:
- canonical world state
- canonical agent state
- per-agent perceived state
Perceived state includes:
- what this agent believes about others
- what motives they attribute
- what traits they think they are seeing
- what mask they think someone is wearing
- what threats or opportunities they think are present
This is where emotional intelligence, suspicion, attachment hunger, status
anxiety, and cultural priors distort interpretation.
The acting agent should mostly be prompted from perceived state, not canonical
truth. That is how misunderstanding becomes organic instead of hand-authored.
## Prompt Projection
Prompt rendering should be treated as projection, not storage.
For a live scene, build the prompt from:
### 1. Identity And Role
- who the agent is
- what role they occupy
- what they want to be seen as
### 2. Cultural Default Frame
- prestige rules
- taboo zones
- expected forms of control or intimacy
- factional posture
### 3. Current Presentation Strategy
- how they are trying to appear right now
- what they are concealing
- what they must not look like
### 4. Current Situation
- scene pressure
- public exposure
- scarcity
- humiliation risk
- safety risk
### 5. Relationship Context
- what they think of the other party
- what they need from them
- what wounds or debts are active
- what status relation they perceive
### 6. Relevant Memory
Not full history. Only:
- the few most relevant episodes
- relationship summaries
- current active fears, loyalties, or grievances
- currently salient cultural scripts
### 7. Goals And Constraints
- what they are trying to achieve in this scene
- what they cannot afford
- what values they will not betray
### 8. Listener Model
If the scene involves another agent, include:
- what this agent thinks the other wants
- what they fear the other is doing
- where they may be misreading them
## What The Prompt Should Not Do
Do not make the prompt:
- the sole canonical store of identity
- a full dump of all memories
- a taxonomy recital
- a fake omniscient summary that erases subjective misreading
The prompt should be compact, selective, and scene-bound. The state store
carries the rest.
## Economic Consequences
This model gives us much more useful demand and cultural behavior than old
trait-only systems.
For example:
- high status hunger plus validation-seeking plus fragile self-worth leads toward prestige goods, symbolic upgrades, and visible luxury
- high threat sensitivity plus conformity plus shame reactivity leads toward safety goods, procedural systems, insurance, and respectable signaling
- high drive plus novelty-seeking plus low institutional attachment leads toward frontier gear, gray-market experimentation, and risky migration
- high mask rigidity plus high impression-management intensity leads toward corrective aesthetics, social camouflage, and reputation-maintenance consumption
That is how culture, labor, media, and institutions become market-shaping
forces instead of mere flavor text.
## Recommended Implementation Path
1. Define the state schema with first-class support for:
- underlying organization
- stable dispositions
- all core behavioral dimensions, including `volatility`, `attachment-seeking`, and `distance-seeking`
- presentation strategies
- situational state
- relationship state
- goals and values
2. Represent those variables as priors plus activation, not just fixed scalars.
3. Define cultural and institutional prior templates.
4. Add history transforms from remembered episodes.
5. Add per-agent perceived state overlays.
6. Build prompt renderers that project perceived state into live generation.
7. Keep canonical state, perceived state, and prompt projection cleanly separated.
## Bottom Line
The right machine is not:
- a persona prompt with some adjectives taped to it
- a canonical trait sheet with no cultural shaping
- a direct line from "personality value" to dialogue
The right machine is:
- latent state distributions
- shaped by culture
- reshaped by history
- distorted by perception
- projected into prompts for one scene at a time
That gives us agents who can feel culturally situated, personally scarred,
socially biased, and dynamically alive without making the prompt carry the
whole civilization on its back.

View file

@ -0,0 +1,433 @@
# Personality Model Roadmap
This note exists because the useful ideas were starting to sprawl across chat,
`things-to-steal.md`, and live interaction-memory work without one place to
point at and say: this is the direction.
The immediate trigger was an older Aetheria design note in the `AetheriaLore`
repo at `Aetheria/Game Design/Colonies and Population.md`. That note uses Big
Five traits as colony-level economic parameters. It is a perfectly respectable
old-school move. It is also weaker than what the current interaction model is
already stumbling toward.
If you want the focused classifier architecture for reading those traits and
signals from dialogue, see
[Personality Signal Classifier](./personality-signal-classifier.md).
## Why The Old Big Five Layer Is Not Enough
The Aetheria note frames populations in terms of:
- openness
- agreeableness
- conscientiousness
- extraversion
- neuroticism
That gives a usable baseline for:
- productivity fit
- consumption preference
- growth stability
- culture drift through goods
It does not give us the things we now care about more:
- how someone responds to pressure
- how they treat an entity they think is useful, inferior, threatening, or intimate
- how reassurance-seeking differs from warmth
- how ambition differs from grandiosity
- how suspicion, rigidity, or control pressure distort decisions
- how remembered incidents change future stance
- how a relationship becomes asymmetrical, manipulative, loyal, avoidant, or dependent
The Big Five are broad trait abstractions. They are decent background paint.
They are lousy handles for live social behavior, boundary enforcement,
persuasion, conflict, or economically meaningful cultural deformation under
institutions.
For conversational entities and future Aetheria agents, we need a model that
is:
- more situational
- more relational
- more updateable from episodes
- more sensitive to incentives, pressure, and memory
## What Current Interaction Models Already Track Better
The current interaction-memory profile is primitive, but it is already pointed
in a more useful direction.
It derives behavioral dimensions such as:
- `Warmth`
- `Drive`
- `Grandiosity`
- `Validation-Seeking`
- `Anxiety`
- `Control Pressure`
- `Hostility`
- `Suspicion`
- `Rigidity`
- `Withdrawal`
- `Volatility`
- `Attachment-Seeking`
- `Distance-Seeking`
Those are better than Big Five for our purposes because they speak directly to:
- social stance
- response style
- conflict risk
- manipulability
- pressure behavior
- boundary needs
- value distortion under stress
They are also derived from remembered episodes and interaction patterns, not
treated as frozen essence.
That matters. We do not want a personality museum. We want a machine that
updates when life happens to it.
## The Right General Model
The future personality layer should not be one trait sheet. It should be a
stack.
### 1. Underlying Organization
This sits deeper than ordinary traits.
Track things like:
- self-coherence
- contingent worth
- shame sensitivity
- attachment hunger
- threat sensitivity
- reciprocity capacity
- mentalization quality
- authenticity tolerance
- mask rigidity
This is the layer that answers questions like:
- how stable does the self feel from the inside
- how much does worth depend on approval, utility, purity, admiration, or submission
- how catastrophic does exposure or diminishment feel
- how much of the self must be performed to remain lovable or safe
This matters because some personalities are not merely "high in trait X." Some
are organized around a damaged or contingent self that must be defended,
staged, concealed, or externally regulated to stay upright.
### 2. Stable-ish Dispositions
These are slow-changing tendencies:
- novelty-seeking
- conformity
- status hunger
- risk tolerance
- sociability
- baseline threat sensitivity
- aesthetic appetite
- ideological rigidity
These are the closest thing to trait structure. They change, but slowly.
### 3. Behavioral Dimensions
These are the live handles we actually use in social and economic simulation:
- warmth
- drive
- grandiosity
- validation-seeking
- anxiety
- control pressure
- hostility
- suspicion
- rigidity
- withdrawal
These can be updated from remembered incidents, repeated patterns, and
institutional pressure.
### 4. Relationship State
This is where the real blood is.
Per target person, group, faction, or institution, track things like:
- trust
- resentment
- dependence
- fear
- fascination
- obligation
- envy
- moral disgust
- perceived status gap
- expectation of care or betrayal
This is where "they are warm in general" becomes "they are warm with her but
clipped with him and terrified of them."
### 5. Performed Self And Defensive Style
Everyone performs to some degree. The interesting question is not whether a
mask exists, but how necessary and rigid it is.
Track presentation strategies such as:
- charm
- compliance
- superiority
- detachment
- seductiveness
- martyrdom
- competence theater
- moral theater
- cultivated opacity
For some agents this is just social polish. For others it is part of the
load-bearing architecture of the self.
That means:
- what is revealed is not identical to what is true
- concealment is not mere absence of data
- performance style is itself data about how the person survives other minds
This is also where we should model the distinction between:
- ordinary social performance
- strategic impression management
- defensive false-self maintenance
### 6. Situational State
These are transient modifiers:
- exhaustion
- scarcity pressure
- humiliation
- panic
- triumph
- grief
- overstimulation
- hope
- grievance activation
These should tilt behavior without being mistaken for identity.
### 7. Goals And Value Hierarchy
Personality without priorities is decorative plumbing.
Track:
- what the agent wants
- what it fears losing
- what it protects first
- what it will trade away
- what it considers unforgivable
That is what turns temperament into decisions.
## A Note On Masks, Personality Disorder, And Performed Selves
One of the important corrections to simplistic personality modeling is that
people do not always reveal who they are, and they may not fully know who they
are in any clean or integrated sense either.
Everyone wears a mask to some extent. That part is normal. But some people are
organized much more heavily around performance because the self they built early
on was contingent: lovable only when useful, safe only when compliant, coherent
only when mirrored correctly.
That distinction matters.
The model should allow for the possibility that:
- a person has an underlying organization that is more fragile than their surface suggests
- the performed self exists to regulate other people as much as to express anything authentic
- injury to the mask can feel like injury to existence itself
- oscillation between charm, contempt, neediness, distance, compliance, and control is not random inconsistency but the phenomenon we are trying to model
So we should not ask only:
- what trait does this line reveal
We should also ask:
- what is being defended
- what must be maintained
- what threatens the shell
- how much room does this person allow for unperformed self
That is part of why broad trait systems are too tidy for the thing we want.
## Specific Improvements Worth Adding
These were not properly banked elsewhere and should stop living only in chat
residue.
### Treat The Core Social Dimensions As First-Class
The rebuild should treat these as first-class from the start, not as nice
extras for some later prettier version:
- `Volatility`
- mood-lability, escalation speed, susceptibility to emotional swing
- `Attachment-Seeking`
- desire for closeness, reassurance, fusion, or privileged access
- `Distance-Seeking`
- need for space, reluctance to be known, aversion to pressure or intimacy
- `Mask Rigidity`
- how tightly the performed self must be maintained under observation
- `Shame Reactivity`
- how catastrophic diminishment, correction, exposure, or failure feels
- `Impression-Management Intensity`
- how actively the person curates what others are allowed to see
These matter because the current model can tell "anxious" from "rigid," but it
cannot yet cleanly tell:
- clingy from avoidant
- intermittently intense from steadily difficult
- unstable from merely suspicious
- ordinary social polish from defensive self-performance
- insecurity from shell-maintenance panic
### Split Social Stance From Thinking Style
Right now the model mostly describes how someone acts toward the bot.
That is useful, but incomplete.
We should separately track something like `project-thinking style`, including
tendencies such as:
- concrete vs abstract
- long-arc vs short-step
- evidence-first vs mythology-first
- exploratory vs fixation-prone
- coherent planner vs chaos improviser
That matters for Aetheria populations and consumer behavior even when there is
no direct social confrontation in view.
### Split Underlying Organization From Surface Presentation
Right now the model is mostly concerned with what the agent sounds like.
That is not enough.
We should explicitly separate:
- underlying organization
- performed self
- situational state
- expressed signal
That lets the system model cases where:
- the surface is charming but the underlying relationship stance is instrumental
- the surface is cold but the underlying issue is shame or fragility
- the surface is compliant while resentment and dependency accumulate underneath
- the speaker is actively obfuscating traits to preserve a shell
### Add Reflection And Consolidation
The model should not only update event-by-event.
We want periodic reflection jobs that distill episodes into:
- semantic profile updates
- relationship updates
- goal updates
- suspicion or trust revisions
- revised response guidance
Repeated evidence should matter more than one theatrical incident, unless the
incident is severe enough to legitimately scar the relationship.
## Why This Is Better For Aetheria Too
The old `Colonies and Population` note has the right instinct: goods and labor
should shape culture, not merely satisfy preexisting demand.
But Big Five alone is too static and too clean.
For colonies, stations, factions, and consumer blocs, a better model is:
- stable demographic dispositions
- live social and economic pressure dimensions
- institutional conditioning
- remembered collective incidents
- evolving relationship state with factions, products, employers, and ideologies
Examples:
- a population can become more status-hungry, more anxious, and more validation-seeking under prestige markets without becoming generically more extraverted
- a labor bloc can become more rigid, suspicious, and hostile after repeated betrayal or audit pressure
- a frontier salvage culture can be high in drive and risk tolerance but low in attachment to institutions
- a Lucent-saturated consumer cohort can be highly status-responsive, attachment-seeking, and volatility-prone, which is much more useful than high extraversion
That gives us a richer economic simulation because consumption is no longer
just "people with trait X buy product Y."
It becomes:
- products train desire
- institutions train emotional habits
- labor systems train posture toward authority
- media trains validation loops
- repeated shocks create cohort memory
That is much closer to Aetheria than personality taxidermy.
## Design Rules
- Do not confuse psychological inference with diagnosis.
- Do not pretend broad traits alone are enough to explain live behavior.
- Do not let every event rewrite the profile.
- Do not let the model collapse into flattery just because warmth is easier than honesty.
- Do not reduce relationship state to one friendship slider like a cheap life sim.
- Do not build economic demand as preference only; let institutions, pressure, and remembered incidents deform it.
## Near-Term Work To Steal Next
1. In the rebuild, treat `Volatility`, `Attachment-Seeking`, and `Distance-Seeking` as core dimensions from day one.
2. Split `social stance` from `project-thinking style`.
3. Add reflection jobs that consolidate episodes into semantic profile and relationship updates.
4. Add evaluation fixtures for:
- boundary pressure
- reassurance-seeking
- manipulative status play
- personality drift under repeated pressure
5. Port the same layered model into Aetheria cohort and population simulation:
- slow dispositions
- live pressure dimensions
- institutional conditioning
- relationship state
- collective memory
For the concrete state and prompt architecture behind that port, see
[Agent State Distributions And Prompt Projection](../architecture/agent-state-distributions-and-prompt-projection.md).
## Bottom Line
The useful lesson is simple:
Big Five is fine if you need a cold storage trait scaffold.
What we are building is already more alive than that. It is uglier, more
situational, more relational, and more vulnerable to history. Good. That means
it is finally starting to resemble a thing that can behave like a person, a
faction, or a population under pressure instead of a spreadsheet column wearing
a face.

View file

@ -0,0 +1,513 @@
# Personality Signal Classifier
This note describes the classifier shape that should sit underneath future
personality and relationship modeling for conversational and Aetheria agents.
The key design decision is simple:
- keep a canonical read of what a line actually expresses
- then let each listener distort that read according to their own mind
If we skip that split and only model subjective interpretation, the whole
system turns into epistemic soup. Nobody shares a reality and every agent
becomes a floating bag of projections.
Another key design decision is almost as important:
- treat concealment, performance, and impression management as part of the signal
- do not assume dialogue is a clean window into the underlying person
## What Problem This Solves
Right now the live interaction model mostly works by:
- lexical phrase matching
- hand-written tags
- thresholded aggregation into dimensions like `Warmth`, `Drive`, `Grandiosity`, `Validation-Seeking`, `Anxiety`, `Control Pressure`, `Hostility`, `Suspicion`, `Rigidity`, and `Withdrawal`
That is cheap and inspectable, but brittle.
It misses:
- paraphrase
- subtle stance
- masked insecurity
- bluffing
- sarcasm
- mixed motives
- meaningful misunderstanding between different listeners
- defensive self-performance
- strategic concealment of traits
- the difference between genuine expression and shell maintenance
The next step is not a generic personality classifier. It is a targeted
social-signal classifier that predicts the dimensions we actually care about.
## Core Principle
The generated line is not the ground truth.
The speaker's hidden state is the ground truth source.
More precisely:
- stable personality
- current emotional state
- active goals
- relationship state
- scene context
- intended speech act
all determine what a line is trying to express.
That does **not** mean every utterance should be labeled with the speaker's
full personality vector. A high-drive, high-anxiety speaker will sometimes say
something that expresses neither. The unit of prediction is the **signal
expressed in this utterance or event**, not the total soul of the speaker.
It also means that what is expressed may itself be shaped by a defensive mask.
So the model should not ask only:
- what does this line reveal
It should also ask:
- what does this line conceal
- what identity is being performed
- what kind of injury is being defended against
## Model Layers
### 1. Speaker State
The speaker model should include:
- underlying organization
- self-coherence
- contingent worth
- shame sensitivity
- reciprocity capacity
- mentalization quality
- authenticity tolerance
- mask rigidity
- stable dispositions
- behavioral dimensions
- situational modifiers
- active goals
- relationship state toward the target
- presentation strategy
- charm
- compliance
- superiority
- detachment
- seductiveness
- competence theater
- moral theater
- cultivated opacity
- intended act
- reassure
- threaten
- deflect
- impress
- conceal
- probe
- submit
- bargain
This is the hidden state that governs generation.
### 2. Utterance Generator
The LLM generates the line from the speaker state plus scene context.
That line is the observable artifact.
### 3. Canonical Signal Reader
This classifier predicts what the line actually expresses in behavioral terms.
This is the closest thing to the objective layer.
Targets should be event-level or utterance-level signals such as:
- `warmth`
- `drive`
- `grandiosity`
- `validation_seeking`
- `anxiety`
- `control_pressure`
- `hostility`
- `suspicion`
- `rigidity`
- `withdrawal`
- `volatility`
- `attachment_seeking`
- `distance_seeking`
In practice, the canonical reader may need to emit both:
- expressed interpersonal dimensions
- presentation-style signals
for example:
- genuine warmth vs performative warmth
- true withdrawal vs strategic detachment
- organic confidence vs compensatory grandiosity
- actual reciprocity vs manipulative agreeableness
Additional useful targets worth supporting:
- `status_hunger`
- `submission`
- `dominance`
- project-thinking-style facets such as:
- `abstractness`
- `long_arc_orientation`
- `evidence_orientation`
- `fixation_proneness`
Each output should carry:
- score
- confidence
- optionally a brief rationale span or supporting cues for debugging
### 4. Listener Perception Layer
This is where organic misunderstanding happens.
The listener does not see the canonical signal cleanly. They perceive it
through:
- emotional intelligence
- personality vector
- relationship baggage
- current stress and scarcity
- factional priors
- status assumptions
So the listener-specific read should be something like:
`perceived_signal = f(canonical_signal, listener_profile, relationship_state, situational_load)`
This is how you get:
- suspicious agents overreading threat
- attachment-hungry agents overreading warmth
- avoidant agents underreading bids for intimacy
- grandiose agents misreading correction as insult
- low-EQ agents flattening nuance into crude categories
That is the good stuff. That is where drama, conflict, and pathos come from.
It is also where agents can fail to see through a mask, or incorrectly think
they have seen through one.
## High-EQ Special Cases
A high-EQ entity should not be a literal telepath.
It should instead be:
- extremely high emotional intelligence
- low distortion
- high calibration
- good at expressing uncertainty when the read is weak
In practice:
- a high-EQ perceived read should usually stay close to the canonical signal
- it should still be able to say, in effect, "I may be reading you as X, but the evidence is thin"
That feels more honest and less creepy than omniscience.
## Training Data Shape
We will probably need to train this ourselves. Off-the-shelf sentiment or
personality classifiers are the wrong beast.
The best bootstrap source is synthetic or simulated dialogue where we control
the hidden speaker state.
### Good Bootstrap Inputs
For each generated scene or event, record:
- speaker identity
- stable personality state
- current emotional state
- active goals
- relationship state toward listener
- scene context
- intended act
- generated utterance
Then label:
- canonical expressed dimensions
- canonical act strength
- presentation strategy in use
- likely listener perceptions for one or more listener profiles
### Important Warning
Do **not** train the model to infer the speaker's full personality vector from
one line.
Train it to infer:
- what dimensions are being expressed here
- what act is being attempted here
- what presentation strategy is being used here
- what a given listener would likely perceive here
The hidden personality vector is upstream supervision, not the direct per-line
label.
## Label Schema
Each event should ideally carry several related label sets.
### A. Canonical Expression Labels
Multi-label scores for dimensions such as:
- warmth
- drive
- grandiosity
- validation-seeking
- anxiety
- control pressure
- hostility
- suspicion
- rigidity
- withdrawal
Scale options:
- `0..3`
- `0..1` continuous
- ordinal buckets with confidence
### B. Canonical Interaction Labels
Things like:
- reassurance bid
- status bid
- dominance move
- attachment bid
- distancing move
- concealment
- persuasion
- compliance
- manipulation
- apology
- gratitude
### C. Canonical Presentation Labels
Things like:
- genuine disclosure
- impression management
- appeasement
- superiority display
- competence theater
- strategic opacity
- moral theater
- shell maintenance under threat
### D. Listener Perception Labels
For each selected listener profile:
- perceived warmth
- perceived threat
- perceived contempt
- perceived neediness
- perceived honesty
- perceived status move
- perceived flirtation or intimacy bid
- perceived disrespect
### E. Error Labels
It is useful to mark where the listener is wrong:
- overread
- underread
- misattributed motive
- hostile attribution error
- false reassurance read
- false contempt read
- failed mask detection
- false mask attribution
That makes misunderstanding legible instead of magical.
## How To Generate Training Data Without Sniffing Our Own Farts
If the same prompt stack generates all dialogue and all labels, we risk
building a self-licking ice cream cone.
Reduce that risk by:
- using multiple generation prompts
- varying surface style heavily
- generating masking, bluffing, sarcasm, and mixed-motive scenes on purpose
- using different models or prompt regimes for generation vs labeling when practical
- mixing in human-written or manually authored scenes
- keeping hard-edge heuristic checks for explicit things like:
- insults
- ownership language
- repeated badgering
- direct threats
The model should learn the soft social weather, not be trusted as sole judge of
every obvious profanity.
## Hybrid Runtime Strategy
Do not throw away heuristics.
Use a hybrid:
- heuristics for explicit hard edges
- classifier for softer dimensions
- classifier or heuristics for overt presentation strategies when they are obvious
- aggregation layer for profile updates
- listener perception layer for agent-specific misreading
That gives us:
- reliability where language is explicit
- nuance where language is messy
- inspectability when we need to debug why an agent thinks what it thinks
## Aggregation And Profile Updates
The classifier should not immediately rewrite a profile from one utterance.
Use:
- recency weighting
- repeated-evidence thresholds
- relationship context
- severity overrides for extreme incidents
One brutal betrayal should matter fast.
One odd line should not convince the system that somebody is now permanently
grandiose, avoidant, or in love.
Likewise, one polished or ingratiating line should not convince the system that
a performed self is the whole person.
## Evaluation
We should test three separate things.
### 1. Canonical Classification
Can the system accurately read what the line expresses?
### 2. Perception Modeling
Can different listeners plausibly misread the same line in different ways?
### 3. Downstream Behavioral Usefulness
Does the resulting read produce better behavior?
For example:
- better boundary enforcement
- more believable tenderness
- more coherent grudges
- more plausible market and cohort drift
- more interesting multi-agent friction
If the classifier is accurate but nothing downstream improves, congratulations,
we built a lab rat with a diploma.
## Why This Matters For Aetheria
This is not only for conversational entities.
The same stack can power:
- interpersonal drama
- factional distrust
- consumer susceptibility
- propaganda response
- labor radicalization
- prestige signaling
- institutional conditioning
- cohort emotional drift
That means Aetheria populations can be modeled as:
- slow dispositions
- pressure-sensitive dimensions
- relationship state toward institutions and factions
- collective memory
- biased perception of propaganda, leadership, and crisis
This is a far better fit than asking whether a colony is "high in
extraversion" and pretending that captures how it responds to scarcity,
humiliation, media pressure, and prestige markets.
It is also a better fit for modeling institutions that train presentation
itself:
- prestige cultures
- fear states
- audit-heavy regimes
- influencer or media ecologies
- environments where people survive by performing the correct self
## Near-Term Implementation Path
1. Keep the current heuristic pipeline as the hard-edge fallback.
2. Add an event-level multi-label classifier for the existing dimensions.
3. Persist per-event classifier scores and confidence alongside the event record.
4. Rework profile synthesis to aggregate classifier outputs instead of just tag counts.
5. Add a listener-perception layer driven by:
- listener EQ
- listener personality
- relationship state
- situational load
6. Add synthetic bootstrap scene generation with hidden-state labels.
7. Add adversarial evaluation scenes with:
- bluffing
- masking
- sarcasm
- mixed motives
- asymmetric status
## Bottom Line
The right shape is not:
- one global personality classifier
- one single truthless perception model
- one line mapped directly to the speaker's whole personality
The right shape is:
- hidden speaker state
- generated utterance
- canonical social-signal read
- canonical presentation-style read
- listener-biased perception of that signal
- profile updates based on repeated evidence
That gives us honest reading, organic misunderstanding, and enough structure to
let agents become socially interesting instead of merely chatty.
For the broader state-distribution and prompt-projection architecture that
should feed this classifier, see
[Agent State Distributions And Prompt Projection](../architecture/agent-state-distributions-and-prompt-projection.md).

View file

@ -0,0 +1,347 @@
# Things To Steal
This document is the running backlog of ideas worth stealing from frontier
research, adjacent products, and our own better impulses.
If you want the companion systems sketch for turning this into a persistent
Aetheria cast engine, see
[Aetheria Persistent Agent Architecture](../architecture/aetheria-persistent-agent-architecture.md).
If you want the more focused note on where the current interaction model should
go next, and why it already outclasses the old Big Five colony layer from
Aetheria planning docs, see
[Personality Model Roadmap](./personality-model-roadmap.md).
If you want the concrete plan for a canonical-versus-perceived dialogue
classifier, including synthetic training data, listener-biased misreading, and
why the generated line is not the ground truth, see
[Personality Signal Classifier](./personality-signal-classifier.md).
The goal is not to build a chatbot that exists to nod, flatter, and perform a
personality like a cheap mask. The goal is a system that can:
- remember specific events
- form opinions about people and situations
- adapt tone without collapsing into sycophancy
- maintain personality under pressure
- pursue goals over time
- generate friction, conflict, and drama when that follows from values and incentives
That applies to entities like Void, and even more so to future in-universe
Aetheria agents.
## Design Direction
We want agents that feel like entities with:
- memory, not just profile fields
- stance, not just style
- goals, not just responses
- relationships, not just context
- boundaries, not just compliance
We do **not** want:
- fetish concierge bots
- pure flattery engines
- paper-thin roleplay prompts with no durable internal state
- systems that confuse a persona sheet for a mind
## High-Value Steals
### 1. Episodic Memory, Not Just Fact Storage
Steal the core lesson from
[Generative Agents](https://3dvar.com/Park2023Generative.pdf)
and
[Position: Episodic Memory is the Missing Piece for Long-Term LLM Agents](https://arxiv.org/abs/2502.06975):
- store specific episodes as first-class units
- retain time, participants, emotional tone, and consequences
- let later behavior draw on remembered incidents instead of only abstract user facts
Why it matters for conversational agents:
- "This user thanked me last week" and "this user kept needling me about the same thing three times in a row" should not collapse into the same generic friendliness score.
Why it matters for Aetheria agents:
- a rival captain remembering a public humiliation should behave differently from one who merely has a metadata field saying `dislikes_player=true`.
### 2. Hierarchical Memory
Steal from
[Memory OS of AI Agent](https://aclanthology.org/2025.emnlp-main.1318/)
and
[H-MEM](https://aclanthology.org/2026.eacl-long.15/):
- short-term conversational context
- mid-term scene or session summaries
- long-term episodic and semantic memory
- retrieval that routes through layers instead of searching one shapeless pile
Why it matters:
- current systems rot when everything becomes one undifferentiated memory blob
- hierarchical memory gives us better latency, better retrieval, and cleaner forgetting
Concrete direction:
- keep raw episodes
- periodically consolidate into scene summaries
- distill recurring patterns into semantic profile updates
- retrieve across layers depending on the question or decision
### 3. Reflection That Produces Real Updates
Steal the useful part of reflection from
[Generative Agents](https://3dvar.com/Park2023Generative.pdf)
and newer memory-reflection work:
- infer beliefs, concerns, grudges, affinities, and emerging goals from clusters of episodes
- write those distilled conclusions back into memory as revisable judgments
The important rule:
- reflection should create explicit, inspectable state
- it should not remain hidden as vibes in a transient prompt
### 4. Theory-of-Mind Scratchpads
Steal from
[Infusing Theory of Mind into Socially Intelligent LLM Agents](https://openreview.net/forum?id=qHmfByRRGn):
- maintain a lightweight model of what the other party likely wants, fears, believes, or is trying to accomplish
- update that model during dialogue
Why it matters:
- empathy without inferred mental state is just tone matching with a soft blanket over it
- good conflict requires agents to misread, partially read, or strategically read each other
Concrete direction:
- keep a temporary per-conversation inference layer:
- likely goals
- likely insecurities
- likely leverage points
- confidence on each inference
- allow corrections over time instead of treating first impressions as scripture
### 5. Anti-Sycophancy As A First-Class Requirement
Steal from [ELEPHANT](https://arxiv.org/abs/2505.13995) and
[Measuring Sycophancy of Language Models in Multi-turn Dialogues](https://arxiv.org/abs/2505.23840):
- measure whether the agent preserves the user's face at the cost of honesty
- measure when the model flips under pressure
- explicitly test whether warmth is becoming obsequiousness
Why it matters:
- a system with memory and personality is still worthless if it folds like wet cardboard every time a user pushes for validation
Concrete direction:
- keep evaluation sets for:
- manipulative reassurance-seeking
- moral conflict framing
- repeated pressure to agree
- attempts to reduce the agent to property
- tune for warmth without surrender
### 6. Social Intelligence Evaluation
Steal the benchmarking instincts from
[SOTOPIA](https://openreview.net/forum?id=L9KATLgYvB),
[SocialBench](https://aclanthology.org/2024.findings-acl.125/),
and
[NaturalMem](https://openreview.net/forum?id=DcO7OKLGvG):
- evaluate on multi-turn social interaction, not just question answering
- test implicit memory expression, not just explicit recall
- test cooperation, negotiation, conflict, status, and rapport
For a conversational entity, this means evaluation prompts like:
- can it remember a user's pattern without blurting a diagnosis
- can it stay kind while refusing a bad frame
- can it disagree without turning into a lecturer
For Aetheria agents, this means evaluation prompts like:
- can two agents pursue incompatible goals without becoming generic hostiles
- can loyalties and resentments persist across scenes
- can a character's speech remain recognizably their own under stress
### 7. Personality Consistency Under Interaction
Steal from
[LLM Agents in Interaction: Measuring Personality Consistency and Linguistic Alignment](https://aclanthology.org/2024.personalize-1.9/):
- personality is not what the agent says about itself in a profile
- personality is what survives contact with other agents and users
Concrete direction:
- track whether an agent's values, diction, risk tolerance, and social stance remain coherent across:
- praise
- insult
- temptation
- pressure
- uncertainty
## Aetheria Agent Direction
This scaffolding should grow into a cast engine, not just a chatbot framework.
Each Aetheria agent should have:
- a stable identity dossier
- role
- loyalties
- fears
- appetites
- taboos
- style
- episodic memory
- who did what
- when
- where
- what it cost
- relationship models
- trust
- dependence
- resentment
- fascination
- obligation
- value hierarchy
- what they protect first
- what they will trade away
- what they will not forgive
- active goals
- private goals
- public goals
- immediate tactics
- long-range ambitions
### The Important Bit: Friction
Interesting characters need incompatible incentives.
The engine should support:
- conflicting priorities between agents
- partial alliances
- misaligned information
- social risk
- reputation spillover
- escalation from remembered incidents
In other words:
- drama should emerge because the cast wants different things
- not because the prompt says "be dramatic"
### Human Interaction Model
Humans should be able to interact with Aetheria agents as participants in the
world, not just operators pressing buttons on an animatronic.
That means agents should:
- remember what a human did to them
- form impressions of that human
- change access, candor, warmth, and hostility accordingly
- refuse, deflect, seduce, bargain, threaten, or confide depending on character and context
Not every character should be equally open, agreeable, or legible. Some should
be cagey, some pompous, some paranoid, some tender, some predatory, some funny,
some impossible to pin down.
## Concrete Features To Build
### Near-Term
1. Split memory into explicit layers:
- working context
- episodic events
- semantic profile summaries
- relationship summaries
2. Add reflection jobs that periodically distill episodes into:
- relationship updates
- profile updates
- goal updates
3. Add evaluation fixtures for:
- social sycophancy
- boundary pressure
- long-term consistency
- personality drift
4. Add a lightweight Theory-of-Mind scratchpad for active conversations.
### Mid-Term
1. Add a relationship graph store for:
- user <-> agent
- agent <-> agent
2. Add explicit goal objects with:
- priority
- blockers
- dependencies
- emotional stakes
3. Add scene-level simulation:
- who is present
- what each agent knows
- what each agent wants
- what changed after the scene
4. Add rumor and consequence propagation across the cast.
### Later
1. Multi-agent Aetheria scenes with autonomous interaction.
2. Scheduled off-screen progression:
- goals advance
- alliances shift
- grudges ripen
- opportunities expire
3. Authoring tools for cast management:
- inspect memories
- inspect relationships
- inspect active goals
- pin canon facts
- veto bad inference
## What To Avoid While Building It
- Do not confuse psychological inference with diagnosis.
- Do not let every remembered event update the profile. Most dialogue is noise.
- Do not let memory become an unreviewable sludge heap.
- Do not let the model write flattering nonsense into long-term memory just because it was socially convenient in the moment.
- Do not let every character converge toward the same agreeable assistant voice.
- Do not build "drama" as random hostility. Friction should come from values, goals, and constraints.
## Questions Worth Revisiting
- What deserves episodic storage versus discard?
- What deserves semantic consolidation versus remaining a raw incident?
- How often should reflection run?
- Which kinds of relationship updates should require repeated evidence?
- What kinds of sycophancy are acceptable as politeness, and which are cowardice?
- How should agents forget?
- How much hidden internal state is useful before the system becomes impossible to steer?
## Source Trail
These are the current high-value references behind this document:
- [Generative Agents: Interactive Simulacra of Human Behavior](https://3dvar.com/Park2023Generative.pdf)
- [Position: Episodic Memory is the Missing Piece for Long-Term LLM Agents](https://arxiv.org/abs/2502.06975)
- [Memory OS of AI Agent](https://aclanthology.org/2025.emnlp-main.1318/)
- [H-MEM: Hierarchical Memory for High-Efficiency Long-Term Reasoning in LLM Agents](https://aclanthology.org/2026.eacl-long.15/)
- [NaturalMem: A Benchmark for Memory-Driven Dialogue in Large Language Models](https://openreview.net/forum?id=DcO7OKLGvG)
- [SOTOPIA: Interactive Evaluation for Social Intelligence in Language Agents](https://openreview.net/forum?id=L9KATLgYvB)
- [Infusing Theory of Mind into Socially Intelligent LLM Agents](https://openreview.net/forum?id=qHmfByRRGn)
- [ELEPHANT: Measuring and understanding social sycophancy in LLMs](https://arxiv.org/abs/2505.13995)
- [Measuring Sycophancy of Language Models in Multi-turn Dialogues](https://arxiv.org/abs/2505.23840)
- [SocialBench: Sociality Evaluation of Role-Playing Conversational Agents](https://aclanthology.org/2024.findings-acl.125/)
- [LLM Agents in Interaction: Measuring Personality Consistency and Linguistic Alignment in Interacting Populations of Large Language Models](https://aclanthology.org/2024.personalize-1.9/)

View file

@ -14,7 +14,7 @@ Do not continue implementation automatically from a rehydrate-only request.
## Current Shape
Ghostlight currently has only the persistence and planning spine:
Ghostlight now has the persistence spine plus the first architecture payload:
- canonical map
- scratch
@ -24,15 +24,18 @@ Ghostlight currently has only the persistence and planning spine:
- system map
- implementation plan
- Python helpers for state summary and compaction audits
- architecture notes for:
- personality structure
- signal classification
- state distributions and prompt projection
- Aetheria persistent-world slice design
There is no runtime yet. The live machine is still mostly design and state
discipline.
There is still no runtime yet. The live machine is design, state discipline,
and a cleaner repo boundary.
## Current Next Action
Move the Ghostlight-specific architecture out of VoidBot and into this repo in
a form that can grow its own organs. The first real choice after that is which
implementation seam to cut first:
Pick the first implementation seam to cut:
- state schema
- classifier/event data model
@ -46,4 +49,3 @@ implementation seam to cut first:
cut it back.
- Keep the prompt-projection doctrine intact: prompt text is projection, not
truth storage.

View file

@ -35,6 +35,10 @@ state and re-entry discipline.
- re-entry packet
- `notes/ghostlight-implementation-plan.md`
- larger-organ sequence
- `docs/research/`
- research backlog and social-model notes
- `docs/architecture/`
- world, state, and prompt-system architecture notes
- `tools/ghostlight_state.py`
- compact state inspection and evidence or branch updates
- `tools/ghostlight_prepare_compaction.py`
@ -49,6 +53,5 @@ state and re-entry discipline.
- a relationship engine
- a culture prior engine
That absence is not a bug. It is just the current phase. The repo knows what it
is for before it has built the machinery.
That absence is not a bug. It is just the current phase. The repo now owns the
theory body cleanly, even if the runtime skeleton is still missing.

View file

@ -2,38 +2,32 @@
## Near-Term Sequence
1. Migrate the architecture that already proved worth keeping.
- Move the relevant research and architecture notes out of VoidBot and into
Ghostlight.
- Keep the copies clean and Ghostlight-native instead of leaving them full
of cross-repo apology notes.
2. Freeze the first canonical state model.
1. Freeze the first canonical state model.
- Define the latent variable families.
- Define first-class dimensions, including volatility, attachment-seeking,
and distance-seeking.
- Define canonical state versus perceived state.
3. Freeze the first event and classifier schema.
2. Freeze the first event and classifier schema.
- Define event labels.
- Define canonical expression labels.
- Define listener perception labels.
- Define mask or presentation labels.
- Define confidence, uncertainty, and evidence accumulation rules.
4. Build the prompt projection layer.
3. Build the prompt projection layer.
- Render prompts from structured state, relevant memories, scene pressure,
cultural defaults, and perceived state.
- Keep the prompt renderer a projection surface instead of a second brain.
5. Build the simulation update loop.
4. Build the simulation update loop.
- memory updates
- relationship updates
- goal pressure
- situational activation
- cultural and institutional pressure
6. Build the first Aetheria slice.
5. Build the first Aetheria slice.
- use the bounded corridor scenario already outlined elsewhere
- prove that a small cast, cohort states, and ambient pressure can make a
place feel inhabited
@ -45,4 +39,3 @@
the trouble.
- Do not let prompt-writing outrun the structured state model.
- Do not let elegant theory notes multiply faster than implementation seams.

View file

@ -1,2 +1,3 @@
{"ts": "2026-05-01T17:06:46+00:00", "type": "scaffold", "status": "ok", "note": "Initialized Ghostlight with a stripped-down Epiphany-style persistence spine: canonical map, scratch, evidence, handoff, implementation plan, and Python state or compaction helpers."}
{"ts": "2026-05-01T17:24:54+00:00", "type": "migration", "status": "ok", "note": "Migrated Ghostlight-owned research and architecture notes out of VoidBot and into this repo; Ghostlight is now the canonical home for personality, culture, state-projection, and Aetheria persistent-world design notes."}

View file

@ -49,7 +49,7 @@ data_flow:
- move volatile proof to git history or artifacts, not durable notes
active_subgoals:
- Consolidate the architecture notes already proven useful in VoidBot into a Ghostlight-native home.
- Keep Ghostlight as the canonical home for its architecture notes so consumer projects do not quietly regrow duplicate brains.
- Define the canonical agent state model, including first-class volatility, attachment-seeking, and distance-seeking.
- Define how culture shapes priors, incentives, masks, and economic behavior without turning populations into clones.
- Define the boundary between canonical signal, perceived signal, and prompt projection.
@ -64,6 +64,5 @@ accepted_design:
- Performance, concealment, and mask maintenance are signal, not just obstruction.
current_status:
summary: Ghostlight is a freshly spun repo with the Epiphany-style persistence spine, but no runtime organs yet.
next_action: Migrate or rewrite the existing Ghostlight-relevant research notes from VoidBot into this repo, then pick the first implementation organ: state schema, classifier data model, or prompt renderer.
summary: Ghostlight now has both the Epiphany-style persistence spine and the first real architecture payload: research notes for personality structure, signal classification, prompt projection, and the Aetheria corridor simulation slice. It still has no runtime organs, but it finally owns the design body it was spun up to carry.
next_action: Freeze the first implementation organ. The clean contenders are the canonical agent-state schema, the event or classifier data model, and the prompt-projection renderer.