Research behind CRI
We study how individual experience becomes persistent state: memory, belief, expectation and relationship, and how new experience can revise what a character carries forward. The working-paper summary sets out the architecture; the notes follow its development inside Luminids.
Working paper
LuminidsAI: A Shared Event-Substrate Architecture for Interpretable Developmental Artificial Life in Games
Abstract, opening
LuminidsAI is a developmental artificial-life architecture for real-time games in which lived events provide a shared causal substrate for memory, awareness, personality, social belief, culture, action, and player-facing explanation.
01 The position
Our approach begins with how an individual interprets an event, then follows what changes in its beliefs, expectations and behaviour.
We call this ordering emotional-first. The character's state changes before language expresses that change. The research asks whether those connections make a character more coherent and easier for a player to understand.
Each link is a research question: what did the character notice, why did it matter, and what changed because of it?
What this character can see or hear.
How the character weighs the moment.
What it now holds true, on the evidence it has.
What it therefore predicts will happen next.
What it would choose while holding that expectation. Belief does not drive this step yet.
What changes in it once the prediction fails.
02 Four positions
One event should shape memory, awareness, personality, beliefs, relationships and later action through a consistent account of what happened.
Characters should hold beliefs backed by evidence rather than read every relevant truth straight from the simulation.
Shared history should stay one record, while each individual develops its own view of the relationship.
Believable behaviour, clear explanations and reliable saving have to work together within the time a game has to think.
03 Measurement
We propose seven measures, grouped around the player, the character and the developer. Each tests a different part of the same experience.
Does the character read as someone rather than something?
Can a player tell characters apart and predict them?
Do different histories produce different characters?
Does what two characters hold about each other stay coherent through ordinary play?
Does the explanation shown name the real cause?
Is a character the same after loading as before saving?
What does a population cost per frame?
04 Notes
The founder's notes follow architectural decisions, engineering failures and discoveries made by watching real play.
A ninety-minute observation caught behaviour that a passing test suite had missed. A single recursive search explained a severe performance problem. These accounts show how the work changes when it meets a running game.
05 Ten principles
A character needs its own experience, encounters that matter, and behaviour whose causes can be understood.
What reaches a character becomes part of its own history.
A shared moment can change each character differently.
A decision should lead back to its causes.
01 Point of view
The world can contain a fact without every character knowing it. Knowledge must come from perception, memory, inference or communication.
02 Experience
A persistent character should become different because of what has happened to them. Personality is not only configuration.
03 Memory
Stored information is not meaningful memory simply because it can be retrieved.
If an experience matters, it should be capable of changing what the character expects or does later.
04 Relationship
Astra's relationship with Quill is not necessarily Quill's relationship with Astra.
A to B and B to A are allowed to disagree.
05 Fallibility
Artificial characters should be capable of misunderstanding situations, holding stale expectations and forming incorrect beliefs.
New evidence should be capable of changing what they believe.
06 Emotion
Emotional state should not exist only so a character can announce a mood. Appraisal, how the character weighs a moment, should alter attention, expectations, goals, choices, memory and future behaviour.
Emotional-first, the character's state first and its words after, does not mean emotion replaces reasoning. It means events acquire personal meaning before they become behaviour.
07 Causality
If we say a belief caused an action, changing that belief while holding the relevant world state constant should be capable of changing the action.
Otherwise the belief is only decoration.
08 Explanation
The system should not invent a plausible reason after an action has already been chosen.
Explanation should point back into the evidence, beliefs, appraisals and goals that took part in the decision.
09 Language
Language models should give artificial characters language, nuance and expression without becoming the sole authority over who that character is.
The persistent cognitive state belongs underneath the expression layer.
10 Proof
Artificial life should survive outside a carefully prepared demonstration: through time, changing relationships, mistakes, population scale and real players.
That is why we are building it inside Luminids.
1 of 10
