AI Research & Product Lab

Every Action Changes
Who We Are and How We Relate

MindOrigin builds consequence-aware AI systems based on P/G Theory — a framework for modeling how actions transform entity states (P), relationships (G), groups, and long-term behavior.

Our first product, Eden, applies this theory to living characters — game NPCs today, human-aware robot companions ahead — whose trust, memory, emotions, and decisions evolve from what people actually do. Games are the first proving ground; the broader goal is to extend the same relational consequence layer into future products for AI agents, simulations, education, institutions, and society-scale systems.

See Eden in Action Read the Research
What We Do

We build the relational consequence layer for AI agents.

Planning, reinforcement learning, and simulation already track outcomes. What modern AI rarely carries is an explicit, persistent representation of how actions reshape individuals and relationships over time. P/G Theory is that layer: a small fixed core plus optional layers, kept strictly separate from the world model.

Property States (P)

What an entity is, has, or risks losing — health, safety, resources, capability, identity, reputation. Dynamic: it grows, depreciates, and accumulates.

Relationship Fields (G)

How entities are situated toward each other — trust, obligation, threat, kinship, shared identity. Directed, dynamic, and it decides whose P counts.

Affect & Decision

Effective property change — self, plus others weighted by identification — generates primitive affect and ranks candidate actions. Named emotions are an interpretation layer above.

External World Model
Impact Estimator
P/G Update
Affect & Decision

The world model supplies facts; P/G estimates only how those facts change P and G. The boundary is strict — the core stays small while the world grows.

Not a replacement — a companion layer

LLM + P/G, side by side

The LLM provides language, knowledge, and reasoning. P/G provides relational state, behavioral consequence, and long-term social memory. An LLM can propose actions and dialogue; the P/G layer estimates their relational consequences, updates state, and guides selection.

One core, widening circles
Characters

Game NPCs and robot companions with persistent trust, grudges, and gratitude — starting with Eden.

Manners

Etiquette as P/G regulation — small displayed-P costs that stabilize relationships.

Institutions

Trust, legitimacy, and contract reliability as dynamic group fields.

Societies

Reputation propagation, cohesion, and polarization in multi-agent systems.

Research

Papers & Preprints

P/G Theory is an ongoing research program. Start with the background document; the full papers are shared with approved reviewers and partners — request access below.

Start Here Background v1.5 · June 2026

P/G Theory Background Document: Definition, Minimal Core, Dynamic P and G, Manners, and NPC Applications

The accessible overview the papers extract from — what P/G is, why both P and G must be dynamic, and how the framework supports manners and NPC systems

MindOrigin Research

Summary

P/G Theory in plain terms: actions transform entity-level property states (P) and relational fields (G), and a minimal core generates primitive affect from effective property change — with G deciding whose changes count. Covers dynamic P and G, threshold-gated decisions, manners as P/G regulation, and NPC applications.

Contribution: the accessible overview the formal papers build on — read this first.

Request the Document

Access is reviewed individually.

Preprint Foundational 2026

P/G Theory: A Computational Framework for Relational Impact Modeling and Manner Formation in AI Agents

Revised Technical Preprint Draft with Concrete Impact-Estimator Training

mindorigin.io

Summary

The formal framework: actions as transformations over P and G, relational ethical value as a bounded function of predicted changes under explicitly separated hard constraints, and manner as relationship-sensitive policy stabilization. Includes a concrete training pipeline for the relational impact estimator — label structure, training sources, and uncertainty calibration.

Contribution: the foundation of P/G Theory, positioned as a representation layer complementary to alignment, RLHF, and multi-agent learning — not a replacement for them.

Request the Full Paper

Access is reviewed individually.

Preprint Extension May 2026

P/G Valuation: A Relational Representation Layer for Behavioral Diversity and Life-Compatible AI

Extends the P/G framework with the Behavioral Diversity Principle and life-compatible safety conditions

mindorigin.io

Summary

Proposes P/G valuation as a relational representation layer and states the Behavioral Diversity Principle: under the same external conditions, different agents choose different actions because their P/G states, weights, memories, and time horizons differ — covering self-protection, cooperation, etiquette, and sacrifice without one reward label.

Contribution: a descriptive grammar for behavioral diversity with an explicit boundary claim — descriptive power is not normative authority; life-compatible AI still requires hard constraints and human oversight.

Request the Full Paper

Access is reviewed individually.

Preprint Extension May 2026

Minimal P/G Affective Core: Primitive Positive and Negative Affect from Property Change in Relational Agents

A minimal P/G-grounded affective substrate for NPCs, persistent social agents, and multi-agent simulation

mindorigin.io

Summary

A primitive valence mechanism grounded in P/G state change: effective P increase generates positive affect, effective decrease generates negative affect, and G determines whose P matters. Personality, thresholds, habituation, and named emotions are layers above the core.

Contribution: a minimal affective substrate for NPCs and social agents, positioned against OCC, PAD, FAtiMA, ALMA, EMA, and WASABI, with an NPC-oriented implementation protocol.

Request the Full Paper

Access is reviewed individually.

Eden

Living characters beyond fixed scripts.

Eden is a character engine for game NPCs and, ahead, human-aware robot companions that interpret human emotion and respond with empathy. It gives characters persistent social memory — trust, gratitude, fear, and reputation that change future behavior based on what people actually do. Designers author a character in familiar Big Five terms; Eden compiles that personality into P/G-native parameters, makes every mapping inspectable, and verifies behavior in live scenarios.

Eden NPC Personality Editor — core weights, emotion thresholds, response dynamics, decision style, relationship priorities, and live scenario preview for the character Elena Hart

The Eden personality editor — one character, fully inspectable: core weights, emotion thresholds, response dynamics, decision style, relationship priorities, and a live scenario preview. Click to enlarge.

Eden is under active prototype development. The interface shown illustrates the planned authoring, compilation, inspection, and scenario-verification workflow.

1 · Author 2 · Compile 3 · Inspect 4 · Verify
01

Author in Big Five

Designers work with the vocabulary they already know. The Big Five overview is computed and read-only — the single source of truth is the sub-traits. Big Five is not the personality; it is the compiler input.

Personality editor — Elena Hart Modified · 3 changes
Openness0.72computed
Conscient.0.68computed
Extraversion0.55computed
Agreeableness0.83computed
Neuroticism0.28computed
Agreeableness — sub-traits edits recompute overview
Empathy 0.86 → η₂, wP psych
Trust 0.80 → η₁
Forgiveness 0.78 → ID recovery rate
Compiles to 23 P/G-native parameters
02

Compile to P/G parameters

Sub-traits compile into valuation weights, identification coefficients, and dynamics — each with a provenance chip back to the trait that produced it. Uncalibrated values are marked, not committed.

P/G parameters — compiled from authoring layer Read-only
Valuation weights wP
psychological0.31from Empathy 0.86
reputation0.14from Sociability 0.60
safety0.09from Anxiety Sens. 0.25
autonomy0.12from Assertiveness 0.51
Identification coefficients η₁–η₆
η₁1.20Trust 0.80frozen form
η₂1.05Empathy 0.86frozen form
η₅0.90Anxiety Sens. 0.25calibration pending
Expert edits detach a parameter from its source (“manual override”)
03

Inspect every mapping

Click any mapping to see the actual functional relationship, its inputs, and where the output is used in the theory. The tool exposes the real mechanism instead of hiding behavior behind black-box personality labels.

Mapping detail Empathy → η₂ affiliation weight
Compile function
η₂ = η₂₀ × (0.5 + Empathy)
   = 0.77 × (0.5 + 0.86) = 1.05
Base value η₂₀0.77
Current input (Empathy)0.86
Output range0.39 – 1.16
Used in
IDij — identification coefficient (Background v1.5, eq. 23)
ΔPeff aggregation — whose P counts
04

Verify in scenarios

The preview walks the real mechanism — property change under relational modulation, core affect with labeled intensities, and candidate actions ranked by value. The tool demonstrates the theory, live.

Scenario preview — a friend lies to Elena Live
1 · Property change (deception detected)
psychological −0.24 reputation −0.05 others ~ 0
2 · Relation modulation (G = friend)
Trust 0.80 → 0.50  ·  IDij 0.82 → 0.71  ·  ΔPeff amplified under close relation
3 · Core affect activation intensity, 0–1
Concern0.62
Hurt0.44
Anger0.22
4 · Candidate actions, ranked by V
Seek truth calmly, express hurt, ask for honestyV = +0.41
Withdraw and observeV = −0.05
Confront angrilyV = −0.12
Try it — author a personality, watch the character change Interactive
Sub-traits — drag to edit
0.86
0.80
0.25
Compiled parameters
η₁ trust weight1.20
η₂ affiliation weight1.05
η₅ threat weight0.90
wP psychological0.31
Scenario — a friend lies to this character
Relation update
Trust 0.80 → 0.48  ·  IDij 0.81 → 0.67
Core affect
Concern0.62
Hurt0.45
Anger0.24
Actions ranked by V
Seek truth calmly, ask for honesty
Withdraw and observe
Confront angrily
Simplified, illustrative compile functions running in your browser — not the production engine. Forms follow Background v1.5; values are uncalibrated. Try low Empathy with high Anxiety, or low Trust, and watch the ranking flip.

One interaction. Persistent memory.

The same merchant, two different players — and the relationship carries forward into prices, quests, warnings, and reputation.

Player steals from the merchant
Trust 0.40 → −0.14

Refuses trade, raises prices, reports to guards, spreads negative reputation.

Player saves the merchant’s child
Trust 0.40 → 0.86

Discounts, hidden quests, help in danger, support in future conflict.

Why Eden first?

Game characters make P/G consequences visible. A player acts, the relationship changes, the NPC remembers, and future behavior shifts — all observable in minutes. Games give MindOrigin a concrete testbed before the same core expands into robot companions, broader agents, simulation, education, and institutional systems.

NPCs Today · Robots Ahead

One engine, two embodiments

Eden is designed from the start for two kinds of characters: game NPCs and human-aware robot companions. Both run the same P/G core — the same compiled personality, relationship memory, and consequence model — differing only in the world model they read and the actions they can take. In robotics, the challenge is not only moving safely through the physical world, but understanding what actions mean to people: whether an action increases comfort, trust, autonomy, and safety, or stress and perceived threat.

For the robot embodiment, the goal is emotional intelligence people can trust. Eden models human emotional and relational state — comfort, stress, trust, perceived threat — from behavior, dialogue, context, and interaction history, and responds with empathy: choosing actions that protect wellbeing, preserve trust, and respect autonomy. In Eden, empathy is not a slogan; it is a compiled parameter (η₂) that determines whose wellbeing counts in a character’s decisions. A P/G-based companion does not claim to read minds — it earns emotional understanding from interaction.

NPCs come first because games make relational consequences visible in minutes; the robot embodiment follows on the same engine, in the real world.

Numerical values shown are illustrative; parameter forms are frozen while calibration is in progress. Validation experiments are specified in the research program.

Request Eden Demo

We are not just modeling what the world is.
We are modeling what actions mean inside it.

MindOrigin is an independent, Toronto-based AI research and product lab, privately capitalized to develop consequence-aware AI systems. We are building P/G Theory as a general relational representation layer, and Eden as the first product built on top of it.

Get Involved

For game studios: request an Eden demo. For researchers and technical partners: request access to the background document and preprints, or discuss integration. Investors and simulation teams are welcome to reach out directly.

Loading
Your message has been sent. Thank you!