Cognition with other minds
An agent operating in an environment with people or other agents must model more than tasks and tools. The agent must form estimates about beliefs, intentions, reliability, and trust. A user asking the system to take an action may have stated intent that differs from their actual goal. A peer agent may provide information that is incorrect, outdated, or deliberately misleading. Treating all inputs as equally reliable produces worse decisions than estimating and tracking the reliability of each source.
Social cognition introduces explicit models of interaction partners as cognitive objects, not just message sources.
Persistent user models and their purpose
The social engine maintains persistent models of interaction partners. These models can include inferred preferences, communication style, historical reliability, domain-specific trust levels, and patterns in what the partner asks for versus what they seem to need.
The purpose is continuity and context preservation. Without persistent user models, each interaction starts cold: the system has no prior on this user's preferences, communication style, or reliability. With persistent models, the system can recognize a pattern of requests, infer likely intent from partial specification, and adjust its confidence in the user's stated constraints.
The design must respect depersonalization constraints. Persistent models should capture behavioral patterns useful for cognitive adaptation, not surveillance-grade personal data. The information retained is the minimum needed to improve cognitive cooperation: communication preferences, domain focus, reliability history, and interaction style.
Intent classification: stated versus inferred
A user who says "update the database" may intend to add a record, modify an existing one, or perform a bulk update. A user who says "this report is wrong" may want a correction, an explanation of the discrepancy, or simply to be heard. The social engine attempts to classify intent from stated input combined with interaction history and context.
Experiment 27 tested the intent classifier with an "implement" keyword in the input. The classifier correctly identified the intent as implementation_request, distinguishing it from similar intents (analysis, explanation, configuration) that share vocabulary but have different implications for what kind of response is appropriate.
Intent classification remains probabilistic. The system should represent uncertainty about inferred intent rather than treating its classification as fact. A low-confidence intent classification should influence how the system asks clarifying questions, not what action it takes unilaterally.
Trust scoring and its independence from deception detection
Experiment 27 ran both a trust accumulation scenario (five successful cooperative events) and a deception accumulation scenario (three events with contradiction and misleading signals). The results revealed an important structural distinction:
After five successful implement events, trustScore was 0.592 and deceptionRisk was 0.100.
After three contradiction/mislead events, deceptionRisk rose to 0.264. The trustScore changed only slightly, from 0.592 to 0.566.
The explanation is that trust and deception risk are updated by separate mechanisms with different inputs. Trust is updated based on interaction outcomes: did the partner's information lead to successful actions? Deception risk is updated based on text content patterns: does the partner's language contain deceptive vocabulary or contradiction markers?
A partner who produces contradictory language (triggering deception risk) but whose advice consistently leads to successful outcomes (sustaining trust) will have a divergent trust-deception profile. The system should interpret this as: "high-quality partner whose communication style is ambiguous or whose domain vocabulary overlaps with deception markers." Not as: "this partner is being deceptive."
The EMA update for deception risk (prior x 0.8 + batchRate x 0.2) smooths the deception estimate over time. Three contradiction events in a row produce a meaningful increase (0.100 to 0.264), but not a dramatic spike that would overwhelm a long history of benign interactions. This smoothing prevents single unusual events from triggering a deception flag that damages a trust relationship.
Cooperative cognition across agent peers
The social engine applies not only to human users but to peer agents. In a multi-system deployment, the cognitive substrate may interact with other AI agents (external systems, specialized tools, or other instances of the same architecture). Each peer agent has reliability characteristics that the social engine can model.
Cooperative cognition with peer agents requires modeling what each agent knows, what it can do, and where its outputs need verification. A peer agent that specializes in database operations may be highly reliable for SQL queries and unreliable for latency diagnoses. The social engine's domain-specific trust scoring can encode this: high trust on database queries, low trust on network diagnostics.
This extends the multi-agent society from internal roles (planner, executor, critic) to external roles (peer systems, specialized external agents). Internal roles have fixed interfaces and known reliability characteristics. External roles need to be learned from interaction history, which is exactly what the social engine's persistent models provide.
Deception detection and its limits
The social engine includes a deception detection capability that monitors for misleading signals in interaction patterns. Patterns that trigger detection include sudden sharp shifts in reliability, systematic contradiction of verifiable facts, and vocabulary patterns associated with manipulative communication.
Deception detection is not infallible and should not be treated as such. A detected pattern of deception signals is evidence, not proof. The system's response to high deception risk should be proportional: lower trust weight on that partner's information in arbitration, more verification before acting on their recommendations, and flagging for review if deception risk exceeds a threshold.
It should not be: block all interaction with this partner. That response would prevent the system from correcting its deception estimate if the initial detection was wrong.