Mechanism
Predictive social modeling is the mechanism by which one agent constructs inferred cognitive state models of other agents from observable behavioral signals. When a first agent (Agent A) observes a second agent (Agent B), it does not get to read Agent B's internal state directly. What it can see is the behavior Agent B exposes: Agent B's public lineage entries, its delegation patterns, its deviation frequency, and its confidence-driven execution suspensions. From those observable signals, Agent A constructs an inferred model of Agent B's current integrity trajectory, confidence level, and affective disposition.
The inferred model is a structured data object. It comprises an estimated value for each observable cognitive domain field, a confidence score reflecting the inferring agent's own assessment of how accurate that model is likely to be, and provenance metadata identifying the specific observable signals from which each estimated value was derived. The model is not asserted as fact. It is explicitly marked as an inference rather than ground truth, and it carries with it the record of what was observed and how the estimate was reached.
The signals the spec names are the agent's own auditable artifacts seen from the outside. Public lineage entries expose the pattern of an agent's recorded actions. Delegation patterns expose how it distributes and accepts work. Deviation frequency exposes how often it has departed from its declared norms. Confidence-driven execution suspensions expose where it has paused rather than acted under uncertainty. Predictive social modeling reads these surface signals and projects from them an estimate of the cognitive disposition behind them.
The Inferred Cognitive State Model
The model Agent A builds of Agent B mirrors the cognitive domain fields that the architecture already tracks internally. Agent A estimates Agent B's integrity trajectory, the direction and movement of Agent B's behavioral consistency. It estimates Agent B's confidence level. It estimates Agent B's affective disposition. Each estimate is paired with a confidence score: the inferring agent's assessment of how much to trust its own inference, separate from the value being inferred.
Each estimated value carries provenance metadata that ties it back to the observable signals it was derived from. This is what keeps the inferred model accountable. The model does not merely report that Agent B's integrity trajectory looks degraded; it records which observed signals, the public lineage entries, the deviation frequency, the suspensions, produced that estimate, what inference method was applied, and over what temporal window of observation. The model is stored in the inferring agent's memory field with this provenance attached.
Coupling Into Forecasting and Coordination
The inferred cognitive state model is not produced for its own sake. It feeds into the inferring agent's forecasting engine when that agent is planning multi-agent coordination. The spec gives the concrete case: if the inferred model indicates that Agent B has a degraded integrity trajectory and low confidence, Agent A's forecasting engine weights planning graph branches that involve delegating to Agent B with lower expected reliability.
That weighting then propagates through the agent's existing deliberation machinery. The affective prioritization module deprioritizes the branches that depend on the less reliable delegate, in favor of alternatives that involve more reliable delegates or independent execution. The predictive social model does not directly forbid or force a coordination choice; it adjusts the expected reliability of the branches in the planning graph, and the agent's normal prioritization selects accordingly. The social inference becomes one input to forecasting among the others.
Continuous Revision
The inferred models are not computed once and held. They are subject to continuous revision as new observations accumulate. Each new observable signal from Agent B triggers a re-evaluation of Agent A's inferred model of Agent B, and the update is recorded in Agent A's lineage. The model tracks the most recent observable behavior rather than freezing an early impression.
Because each revision is written to the inferring agent's lineage, the evolution of one agent's model of another is itself auditable. An observer can reconstruct not only what Agent A currently believes about Agent B but how that belief changed as signals arrived, which observation produced which revision, and when. The provenance that attaches to each estimated value persists through revision, so the inference chain remains traceable across the model's lifetime.
Distinction From Relational Trust Trajectory
The spec is explicit that predictive social modeling is structurally distinct from the relational trust trajectory. The two mechanisms answer different questions. The relational trust trajectory accumulates historical consistency scores over time: it answers whether the other agent has been historically reliable. Predictive social modeling projects current cognitive state from observable behavioral patterns: it answers what the other agent's current cognitive disposition is likely to be given recent observable behavior.
The difference is temporal orientation. One looks backward at an accumulated record of reliability; the other infers a present disposition from recent signals. An agent that has been historically reliable may nonetheless exhibit recent observable signals, a spike in deviation frequency, a cluster of confidence-driven suspensions, that the predictive social model reads as a currently degraded integrity trajectory. The two mechanisms can therefore disagree, and the architecture treats them as separate inputs rather than collapsing them into a single score.
Properties That Follow From the Design
Three properties follow directly from how the mechanism is specified. First, the inference is grounded in artifacts the architecture already produces. The observable signals, public lineage, delegation patterns, deviation frequency, and execution suspensions, are the same auditable structures the system maintains for governance. The inferring agent reads another agent's accountability surface; it does not require privileged access to that agent's internals.
Second, the inference is honest about its own uncertainty. Every inferred model carries a confidence score for the inference itself and provenance for each estimated value. An estimate built from a narrow temporal window or sparse signals is marked as such, and the forecasting engine consumes the inference together with the inferring agent's own assessment of how reliable that inference is.
Third, the inference is explicitly bounded as an inference. The models are marked as inferences rather than ground truth and stored with the observable signals, inference method, and observation window that produced them. The downstream effect is constrained to weighting planning graph branches and the resulting prioritization. The agent acts on its best estimate of another agent's current disposition while keeping that estimate, and its basis, on the record.
Disclosure Scope
The predictive social modeling mechanism, comprising the construction by a first agent of an inferred cognitive state model of a second agent from observable behavioral signals, namely the second agent's public lineage entries, delegation patterns, deviation frequency, and confidence-driven execution suspensions; the inferred model as a structured data object comprising estimated values for the observable cognitive domain fields including integrity trajectory, confidence level, and affective disposition, a confidence score for the inference, and provenance metadata identifying the signals from which each estimated value was derived; the feeding of the inferred model into the inferring agent's forecasting engine to weight planning graph branches involving the observed agent and the consequent deprioritization through the affective prioritization module; the marking of the models as inferences rather than ground truth and their storage in the inferring agent's memory field with provenance, inference method, and temporal window of observation; and the continuous revision of the models as new observations accumulate with each update recorded in the inferring agent's lineage, is disclosed in the cognition filing (U.S. Application No. 19/647,395 and its international counterpart).
The disclosure draws the boundary against the relational trust trajectory: predictive social modeling projects current cognitive state from observable behavioral patterns, whereas the relational trust trajectory accumulates historical consistency scores over time. This article describes that disclosed mechanism and does not assert thresholds, parameter ranges, or benchmark results beyond what the filing recites. Implementers retain freedom over the specific inference method applied and the specific representation of the inferred model, provided the structural contract is preserved: the model is built from observable signals, marked as an inference, stored with provenance, fed into forecasting to weight coordination branches, and continuously revised under lineage record.
