Other Agent State Model Calibration

Model another agent as having its own partial knowledge, goals, attention, constraints, and interpretations, then update that model from evidence before routing action through it.

Summary

Other-Agent State Model Calibration is the solution pattern for using theory-of-mind reasoning responsibly. It treats another person, role, group representative, counterparty, learner, patient, user, or collaborator as a separate center of knowledge and action. The goal is not to read a mind. The goal is to keep a bounded, evidence-linked, uncertainty-marked model of what the other agent may know, want, fear, attend to, infer, expect, or be able to do, then use that model to choose communication and action that fit the actual interaction.

This draft fills queue position 31 for the accepted prime theory_of_mind. The target definition emphasizes an updateable model of another agent's hidden states. The archetype therefore focuses on the full calibration loop: define whose state is being modeled, separate the actor's own state from the other agent's state, generate alternative hidden-state hypotheses, connect those hypotheses to evidence, filter projection and stereotype, route action through an ethical-use gate, and update after feedback.

Why This Is Not Just “Empathy”

Empathy can be a feeling, a stance, a skill, a performance, or a moral aspiration. This archetype is narrower and more operational. It asks what another agent can know, what they are likely attending to, what constraints they face, what goals or fears may be active, how they may interpret a signal, and how those hypotheses should change after observation. A person can feel sympathetic and still mis-model another agent. Conversely, a calibrated other-agent model can improve care and coordination even when the actor does not share the other agent's emotion.

The important guardrail is humility. Hidden-state claims are hypotheses, not facts. Good theory-of-mind work preserves the other agent's ability to correct the model.

Key Components

Component	Description
Modeled Agent Boundary ↗	The first component names whose hidden state is being modeled and why. A teacher may model a learner's knowledge state; a clinician may model a patient's fear and understanding; a mediator may model each party's perceived risk; a product team may model a first-time user's attention and interpretation. The boundary prevents social inference from expanding into unrestricted psychological profiling.
Actor-State Separation ↗	The actor must distinguish what they know from what the other agent can know. This is the defense against curse-of-knowledge errors. A domain expert, manager, designer, physician, or government official may hold context that is invisible to the recipient. If that context leaks into the simulated perspective, the model becomes a mirror of expertise rather than a model of the other agent.
Hidden-State Hypothesis Set ↗	The archetype works with a set of possible states: beliefs, goals, knowledge gaps, fears, motivations, incentives, attention targets, emotional states, social pressures, and interpretations. Keeping several hypotheses alive prevents premature motive attribution. For example, missed deadlines may indicate overload, confusion, disagreement, lack of authority, low trust, poor incentives, or genuine refusal. Those possibilities call for different actions.
Evidence-Signal Map ↗	Hidden-state hypotheses need evidence. The evidence can be direct statements, questions asked, choices made, hesitation, repeated errors, timing, prior history, social context, or environmental constraints. The map does not eliminate ambiguity; it records ambiguity so later action is less overconfident.
Projection and Stereotype Filter ↗	Other-agent models are vulnerable to two common distortions. Projection makes the other agent too much like the actor. Stereotype makes the other agent too much like a role, group, or label. The filter asks whether the model has become a story the actor already wanted to believe.
Uncertainty and Confidence Annotation ↗	The model should carry confidence levels and alternative explanations. This matters because many observed signals are compatible with more than one hidden state. A quiet learner may be confused, bored, anxious, tired, processing, culturally deferential, or already finished. Good action often has to remain respectful across several of those possibilities.
Action Routing Gate ↗	The model is useful only if it changes action. The gate specifies which decisions may be routed through the model: explanation level, interface wording, negotiation sequence, care plan, support offer, escalation, apology, or question. It also states which uses are forbidden, such as manipulation, profiling, coercion, or unnecessary private-state inference.
Feedback Update Loop ↗	Theory-of-mind reasoning must update. If the other agent reacts unexpectedly, the model should change. Surprise is not proof that the other agent is irrational; it is information that the actor's model may be incomplete or wrong.
Ethical Use Boundary ↗	The same ability that improves coordination can be used to exploit vulnerability. The ethical-use boundary limits inference, retention, sharing, and targeting. It gives special protection to high-stakes contexts, vulnerable agents, dependency relationships, and situations where direct consent or correction is possible.

Common Mechanisms

A Belief-Desire-Knowledge Map documents what the other agent may know, want, believe, lack, fear, or expect. It becomes stronger when each claim is linked to evidence and confidence.

A Perspective-Taking Interview replaces pure inference with direct inquiry. The actor asks open questions about the other agent's understanding, constraints, and priorities, then revises the model.

A False-Belief Check asks whether the actor is accidentally assuming that the other agent knows something they do not know. This is especially useful in teaching, expert communication, onboarding, public instructions, and medical explanation.

An Empathy Map with Evidence Marks can be useful when it records evidence and uncertainty. Without evidence marks, an empathy map is just a worksheet for projection.

A Role-Reversal Simulation can expose blind spots, but it must be checked against evidence. Role reversal based only on imagination often recreates the actor's own assumptions.

A Prediction and Surprise Log records expected reactions, actual reactions, and model updates. This mechanism is valuable in repeated interactions because it makes calibration visible.

A Counterparty Model Red Team challenges a strategic model by generating alternative motives, constraints, and interpretations. This is important in negotiation, diplomacy, organizational conflict, and adversarial coordination.

A Consent and Privacy Boundary Checklist asks whether the hidden-state model is legitimate to create, retain, share, and use. This is crucial when the model concerns health, vulnerability, children, employment, benefits, discipline, or sensitive behavior.

• Active Listening Loop
• Belief-Desire-Knowledge Map
• Consent and Privacy Boundary Checklist
• Counterparty Model Red Team
• Empathy Map with Evidence Marks
• False-Belief Check
• Interaction After-Action Review
• Perspective-Taking Interview
• Prediction and Surprise Log
• Role-Reversal Simulation
• Stakeholder Hidden-Constraint Board

Parameter Dimensions

The archetype varies along several dimensions:

• Agent specificity: a specific person, a role holder, a group representative, or a typical interaction partner.
• State type: knowledge, belief, attention, emotion, goal, incentive, constraint, identity threat, trust, or interpretation.
• Evidence strength: direct report, observed behavior, contextual clue, historical pattern, proxy signal, or weak guess.
• Interaction cadence: one-shot encounter, repeated relationship, continuous service, negotiation sequence, or long-term care.
• Power asymmetry: peer relationship, expert-novice relationship, institution-citizen relationship, caregiver-dependent relationship, or adversarial relationship.
• Ethical risk: low-stakes communication fit, high-stakes support, protected information, manipulative targeting risk, or vulnerable-agent context.
• Updateability: static assumption, episodic update, feedback-rich learning loop, or formal review process.

Invariants to Preserve

The other agent must remain separate from the actor. The model must not smuggle in the actor's knowledge, motives, or values.

Hidden-state claims must remain provisional. The archetype is not successful when it produces a confident story; it is successful when it produces a useful, revisable model.

Uncertainty must remain visible at the point of action. People often behave as if one interpretation is obvious after a story is written down. The artifact should preserve alternatives.

Perspective-taking must not override agency. The other agent should be able to correct the model when possible, and direct communication should be preferred when safe and feasible.

Ethical use must be explicit. Modeling another agent's hidden state can improve dignity and fit, but it can also enable coercion, surveillance, or exploitation.

Neighbor Distinctions

Mental Model Mismatch Repair corrects a person's model of a system. Other-Agent State Model Calibration builds the actor's model of another agent.

Shared Mental Model Alignment creates a common model across a team or group. Other-Agent State Model Calibration can support alignment, but its focal structure is asymmetric: one actor is modeling another agent's hidden state.

Evidence-Grounded Persona Proxy Design creates a reusable proxy for a population. Other-Agent State Model Calibration models a specific agent, role, or counterparty as a separate center of belief and action.

State Estimation infers hidden system state from signals. This archetype is agent-specific: beliefs, goals, knowledge, attention, interpretation, trust, and intention are central.

Credible Signaling designs hard-to-fake signals. This archetype may interpret signals, but its core work is maintaining and updating the other-agent model.

Situational Attribution Check prevents trait over-attribution. This archetype can use that check, but it builds the wider hidden-state model needed for communication, care, teaching, negotiation, or design.

Examples

In education, a teacher checks what a learner can infer before adding a new abstraction. The teacher does not simply ask, “Was that clear?”; they model the learner's knowledge state and test it with a targeted example.

In healthcare, a clinician does not treat hesitation as noncompliance. The clinician models several possibilities: fear, cost, mistrust, misunderstanding, family pressure, side-effect concern, or lack of transportation. The next question tests the model without shaming the patient.

In product design, a team models what first-time users think a button will do. If users treat a destructive action as reversible, the team updates the onboarding and confirmation flow.

In conflict resolution, a mediator tracks each party's perceived threat and dignity needs. A proposal sequence that seems efficient from the mediator's perspective may be unacceptable from a party's hidden-state model unless face-saving or trust repair comes first.

In AI assistant design, the assistant may infer that a user is uncertain, but it should ask clarifying questions and offer control rather than silently profiling the user or exploiting a vulnerability.

Non-Examples

A demographic profile is not this archetype. It may help contextualize a population, but it does not model an agent's hidden beliefs, goals, attention, and update path.

A stereotype about how “people like that” think is not this archetype. It is exactly the kind of ungrounded substitution the archetype is meant to prevent.

A generic machine-state estimator is not this archetype unless the hidden state is agentic and perspective-bound.

A manipulative marketing funnel that infers fear or impulsivity to exploit a user is a misuse, not an implementation.

Review Notes

This draft should remain human-review sensitive because it touches privacy, social inference, power asymmetry, and manipulation risk. The ontology distinction to preserve is that theory_of_mind is not merely mental_model; it is a mental model specifically about another agent's hidden state. It is also not merely shared_mental_model, because the model need not be shared, and not merely state_estimation, because the state is intentional, informational, attentional, and interpretive.

Compression statement

Other-Agent State Model Calibration applies when useful action depends on what another agent can see, believe, infer, intend, value, or tolerate, but those states are hidden, partial, and different from the actor’s own state. The archetype builds a bounded model of the other agent, separates evidence from projection, tracks alternative hidden-state hypotheses with confidence, updates the model through interaction and feedback, and gates use so perspective-taking supports coordination, care, teaching, negotiation, design, or conflict repair without becoming manipulation or unjustified mind-reading.

Canonical formula: agent_boundary + hidden_state_hypotheses + evidence_signal_map + projection_controls + uncertainty_marks + update_loop + ethical_use_gate -> calibrated_action_through_other_agent_model

Related Abstractions

Abstractions this archetype builds on — directly (a source ingredient) or as a related pattern. Links follow the typed catalog namespace.

Built directly on (14)

• Agency: A system pursues representable goals through actions whose selection is sensitive to its beliefs about its situation, via a goal-representation, world-model, and action-selection coupling.
• Attention: The selective allocation of a fixed processing capacity to some inputs while the rest are filtered out, surfacing scarcity upstream of every decision.
• Belief Formation: Commitment-transition by which an agent comes to hold a proposition as true and act accordingly.
• Feedback: Outputs influence inputs.
• Foreseeing (Prediction): Predict future states.
• Hidden Information Reconstruction: An observer combines a system's disclosed outputs with a prior over possible inputs to reconstruct a protected input to whatever resolution the prior permits, so privacy is a property of the (system, adversary-prior) pair rather than of the disclosure surface alone.
• Information Asymmetry: Parties to an interaction hold unequal private knowledge.
• Joint Attention: Two or more agents share an attentional target and each knows the other attends to it.
• Mental Model: Internal system representation.
• Metacognition: Awareness of thinking processes.
• Representation: Model complex ideas.
• Second-Order Cybernetics (Second-Order Observation): Observer within system.
• Theory Of Mind: An agent maintains a separate, updateable model of another agent's hidden states and routes behaviour through that model rather than through ground truth.
• Uncertainty: Incomplete knowledge.

Also references 17 related abstractions

• Consent: Voluntary agreement.
• Curse Of Knowledge: Holding knowledge systematically degrades an agent's ability to simulate the perspective of an agent who lacks it.
• Epistemic Humility: Calibrating the confidence of one's claims to the actual strength of the evidence and staying open to revision when new information arrives.
• Fundamental Attribution Error: Misattribute causes.
• Game-Theoretic Strategy: Strategic interaction analysis.
• Inconsistent Shared Model: Subsystems hold mutually incompatible models of the same external state, undetected until a forcing event requires joint action.
• Object Permanence: A system maintains persistent internal state for entities it cannot currently observe, modeling them as continuing to exist and re-binding them on re-emergence.
• Observational Learning (Social Learning): Learn by observing.
• Perspective: Representation of depth.
• Projection: Map a richer object onto a lower-dimensional target along a chosen direction, discarding the rest.

▸ Show 7 more

Variants

Narrower or domain-specific specializations that share this archetype's core structure. Recognized variants are established; candidate variants are provisional.

Cooperative Perspective-Taking Calibration · communication variant · recognized

A cooperative variant that models what another agent can understand, notice, and safely correct during explanation, teaching, care, support, or conflict repair.

• Distinct from parent: It emphasizes direct clarification, active listening, and shared correction more than strategic inference.
• Use when: The goal is mutual understanding rather than strategic advantage; The other agent can correct or refine the model through dialogue.
• Typical domains: education pedagogy, healthcare delivery, customer support, conflict resolution
• Common mechanisms: perspective taking interview, active listening loop, false belief check

Strategic Counterparty State Modeling · governance variant · recognized

A strategic variant that models another party’s hidden constraints, incentives, beliefs, audience pressures, and likely interpretation of offers or threats.

• Distinct from parent: The parent applies broadly; this variant is tailored to strategic settings where the other agent may shape signals deliberately.
• Use when: Negotiation, diplomacy, bargaining, or adversarial coordination depends on hidden constraints and incentives; Signals may be strategic, incomplete, or costly to verify.
• Typical domains: diplomacy and negotiation, organizational management, public administration policy
• Common mechanisms: counterparty model red team, prediction and surprise log, consent and privacy boundary checklist

Curse-of-Knowledge Countermodeling · affective or cognitive variant · recognized

A variant focused on modeling what a less-informed agent cannot yet know or infer, especially when expertise makes that gap hard to see.

• Distinct from parent: The parent covers other-agent hidden-state modeling generally; this variant centers on knowledge-asymmetry errors created by expertise.
• Use when: Experts explain to novices; A team assumes users, citizens, patients, or learners already know background context.
• Typical domains: education pedagogy, product design, public communication
• Common mechanisms: false belief check, role reversal simulation, active listening loop

Care/Support Needs Inference · domain variant · recognized

A care-oriented variant that models distress, understanding, capacity, fear, and support needs while preserving consent and dignity.

• Distinct from parent: The parent applies across many domains; this variant is tailored to care, teaching, and support contexts where power asymmetry matters.
• Use when: The other agent may struggle to articulate needs or constraints; Support decisions have emotional, developmental, health, or dependency stakes.
• Typical domains: healthcare delivery, caregiving and development, education pedagogy
• Common mechanisms: perspective taking interview, consent and privacy boundary checklist, interaction after action review

Near names: Theory of Mind, Mentalizing, Mindreading, Empathy Map, Stakeholder Empathy.