State, relationships,
and consequences
P/G represents how action consequences change entity properties and relational state, and how those changes contribute to modeled affect and subsequent behavior.
Core representations
What the model makes explicit
P — Entity property state
P represents an entity’s properties, such as health, resources, capability, and reputation. The declared schema specifies which properties an experiment represents and how they change.
G — Directed relationships and group contexts
G includes directed relationships between entities and explicitly defined group-level relational fields. Group-level G is distinct from giving an abstract group its own P-vector.
The current affective baseline uses member identification and descriptive cohesion/continuity summaries. The valuation framework can separately assign decision weights to changes in group-level G.
Primary affect and derived emotions
In the P/G model, effective property changes generate positive or negative primary contributions associated with particular objects. The proposed derived-emotion layer uses personality, current state, and experience to interpret these contributions and associate them with candidate responses.
Relationships, identification, and groups
Relationships are directed: one agent’s relationship toward another need not be reciprocated. In the affective model, an identification coefficient derived from relational state helps determine how another entity’s believed property changes enter the observer’s effective property field.
G also includes explicitly defined group-level relational fields. In the current affective baseline, group context enters through identified members and group summaries describe cohesion and continuity. The broader valuation framework separately includes decision-weighted changes in group-level G; its meta-G proposal discusses commitments concerning humanity, individual protection, ecology, and relations among AI agents.
Group-level G is distinct from assigning an abstract group its own P-vector. Relational state, valuation weights, and protection commitments have separate roles. Relational identification is distinct from moral importance and from final action valuation.
Conceptual architecture
From represented events to candidate responses
Computational verification and mechanistic evaluation of the specified core are complete for the current manuscript cycle. The frozen decision layer and execution constraints have been implemented and computationally tested. Personalized response learning and later personalization layers remain future work. Human correspondence remains untested.
- World information and impact estimation: represent available observations and the consequences assigned or predicted under the declared domain.
- P/G state updates: update property beliefs and directed relational state under specified information and timing rules.
- Primary affect: retain signed contributions with their object and structural origin; their sum is an aggregate report.
- Derived emotions and candidate responses: a proposed personalized layer interprets contributions and associates candidate actions.
- Decision: evaluate feasible actions under a separately specified decision mechanism and execution constraints.
On authored vignettes, a versioned event dictionary supplies consequences. A learned impact estimator is a separate realization; the current studies do not require a trained world model.
Model-scoped predictions
Primary contributions and personal responses
State-dependent intensity
Under the manuscript’s declared normalization and matched conditions, the same resource gain produces a larger positive primary contribution at lower relevant holdings. The prediction applies within the specified range; it is not a claim that lower wealth strengthens every emotion.
A larger primary contribution need not produce a strictly larger reported rating once thresholds and scale ceilings intervene.
Derived emotions and action sets
In the proposed architecture, a derived emotion can organize several candidate actions, and candidate sets can overlap. An emotion does not prescribe one response; observed behavior does not uniquely identify an emotion.
Personalization preserves each primary contribution’s object, sign, and structural origin.
State, personality, and experience
The model distinguishes current state, slower personality parameters, and learned personal history. These have different roles in the proposed interpretation and response layers. Their complete personalized realization requires further evaluation.
Language and explicit state
How P/G relates to language models
P/G specifies state representations and update mechanisms. A language model may provide language capabilities or serve as a comparison system. A proposed language-model integration could use P/G state when evaluating candidate responses and would require its own evaluation. The frozen decision/search layer has been implemented and computationally tested; human correspondence remains untested.
Manners and interaction
Relational consequences beyond immediate costs
P/G examines manners as patterns of interaction involving recognition, attention, access, obligation, status, and their relational consequences. The account extends beyond paying a small courtesy cost to preserve a relationship.
Institutional and society-level applications are future research directions. They require explicit environmental models, group definitions, and evaluation criteria rather than an automatic extrapolation from individual interactions.