Empathy as Distributed Moral Load

Mechanism

The empathy mechanism produces a structured observation: a bounded model of the counterparty's likely affective and intentional state, expressed in canonical fields with explicit confidence bands. The model is built from observable signal — utterances, declared preferences, prior interaction history, contextual classifiers — and never from privileged or unconsented data. Its outputs are consumed by negotiation, harm-evaluation, and tone-adjustment subsystems, but it does not directly write into the agent's policy reference or integrity envelope. The separation of producer and consumer is what makes the mechanism auditable: an observer can inspect what the empathy field claimed and, separately, inspect what the agent's policy did with the claim.

Distinction from behavior cloning is structurally enforced. Behavior-cloning architectures treat the counterparty's expressed preference as a target the agent should approximate; the agent learns to produce outputs the counterparty would produce, or outputs the counterparty would prefer to receive, with no separation between the counterparty's state and the agent's own normative commitments. The empathy mechanism is the opposite: the counterparty model is a sensed quantity, not a target. The agent's policy decides what to do with it, and "do nothing" is always a valid response when the policy does not authorize empathy-driven adjustment for the decision class at hand.

The mechanism distributes moral load by making the counterparty's projected experience a first-class input to harm evaluation. A proposed action is scored not only against the agent's own integrity envelope but against the projected effect on the counterparty's modeled state. Where projected harm is high, the harm-minimization subsystem can require alternative formulations, request consent, or refuse the action. The empathy field is the substrate on which this evaluation runs, and the evaluation itself is structured rather than implicit: the harm score, the inputs that produced it, and the consequence imposed are all separable lineage events.

Influence on policy is bounded. The architecture specifies, in policy, the maximum weight that empathy-derived signals can carry in any decision class. The bound prevents two failure modes: empathy-driven sycophancy, in which the agent abandons its commitments to please the counterparty, and empathy-driven manipulation, in which an adversary feigns affective state to steer the agent's behavior. Both failures arise when empathy signals are unbounded; the bound is therefore a structural safety property, not a tuning knob. An agent operating with empathy bounds correctly enforced cannot be talked out of its principal commitments, no matter how persuasive the counterparty's apparent distress, because the architecture refuses to convert empathy beyond bound into policy adjustment.

The empathy field itself carries lineage. Every observation that feeds the model, every inference produced from it, and every consumption of its output by downstream subsystems is recorded. An auditor can reconstruct what the agent inferred about the counterparty, what confidence the inference carried, and how that inference influenced — within its bound — the action that followed. The lineage permits a regulator to examine, after the fact, whether the agent's empathic inference was reasonable given its inputs, whether the influence the inference exerted respected the bound, and whether the action ultimately taken was consistent with both the inference and the policy.

Asymmetry between sensing and acting is the core architectural commitment. The empathy field can sense richly without that sensing producing rich behavioral consequences. A counterparty in apparent distress will be modeled with high fidelity, but the policy still decides which actions are permissible; the policy may permit gentler tone, or it may not; it may permit consent prompts, or it may not. The richness of empathic sensing is decoupled from the breadth of empathic action by design.

Operating Parameters

Influence bounds are expressed per decision class. A high-stakes class, such as a clinical or financial recommendation, may carry a low empathy-influence bound; the agent's normative commitments dominate, and counterparty affect adjusts only marginal aspects of formulation. A low-stakes class, such as conversational tone in a non-consequential exchange, may carry a higher bound. The class taxonomy is policy-configurable, and the assignment of decision instances to classes is itself a structured operation, not an implicit one buried inside a learned routing layer.

Confidence bands attach to every counterparty-state observation. Where the empathy model has low confidence, downstream consumers receive the observation with that uncertainty made explicit; harm-evaluation must treat low-confidence inferences conservatively rather than acting on them as if certain. The architecture refuses to flatten uncertainty into point estimates, recognizing that a confidently wrong inference about a counterparty's state is more dangerous than an uncertain inference handled with care.

Consent and visibility parameters govern which signals are admissible to the empathy model in the first place. Signals declared by the counterparty are always admissible; ambient observations are admissible only if policy permits, with default settings that bias toward exclusion. The model degrades gracefully when admissible signal is sparse, producing wider confidence bands rather than confabulating; when no admissible signal is available, the empathy field reports an explicit absence rather than a default neutral state, so downstream consumers are not misled into treating no-data as the absence of significant affect.

Decay parameters govern how counterparty-state observations age. Affective states are temporally local; an inference made fifteen minutes ago is not normally evidence about the counterparty's state now. The mechanism decays observations on a policy-configured curve, so that stale inferences carry less weight in current evaluation. Decay curves are configurable per signal type: declared preferences may decay slowly, while inferred mood may decay quickly, reflecting the underlying volatility of each.

Per-counterparty memory parameters govern what is retained across sessions. A long-relationship configuration may permit the agent to retain coarse counterparty-state summaries across encounters; a single-session configuration discards all empathy-derived state at session end. The choice is policy-controlled, visible to the counterparty where required, and never made silently within the model.

Alternative Embodiments

In a negotiation embodiment, the empathy mechanism feeds a bargaining subsystem that uses the counterparty model to identify pareto-improving formulations: phrasings that better fit the counterparty's expressed needs without compromising the agent's principal interests. Influence bounds prevent the bargaining subsystem from conceding the principal's actual position; empathy refines presentation, not substance. The lineage trail in such an embodiment becomes the basis for principal-agent accountability: the principal can verify that the agent did not yield principal value to counterparty pressure.

In a clinical-companion embodiment, the empathy mechanism informs tone, pacing, and disclosure timing for sensitive information. The agent's clinical recommendations remain governed by clinical-governance policy; empathy adjusts how recommendations are delivered, when they are delivered, and what supportive framing accompanies them. The bound prevents empathy from softening a recommendation into ambiguity, so a patient in distress receives the same diagnostic substance as a calm patient, with delivery adjusted but content preserved.

In a multi-counterparty embodiment, such as a mediator agent, the empathy field maintains separate models for each counterparty, with cross-model comparison feeding harm-minimization across the group. The architecture prevents the agent from collapsing distinct counterparties into an aggregated state, preserving the moral salience of each. A mediator that quietly averaged its counterparties' affective states would be less able to recognize which party is being harmed at any moment; the architecture refuses that averaging.

In an adversarial embodiment, where counterparties may feign affect, the influence bound and confidence band together degrade the empathy mechanism's effect on the agent's behavior, allowing the agent to remain functional under manipulation. The mechanism does not need to detect adversarial feigning to be safe against it; the structural bound is sufficient. Detection capabilities, where present, narrow the bound further, but their absence does not expose the agent to unbounded manipulation.

In an educational-tutor embodiment, the empathy mechanism informs scaffolding decisions: when to push, when to ease, when to switch modality. The bound ensures that empathy does not soften the curriculum into vacuity; the learner's apparent distress at a difficult problem is information about pacing, not authority to skip the problem.

In a public-service embodiment, where the counterparty population is large and heterogeneous, the empathy mechanism operates with stricter consent defaults and lower influence bounds, reflecting the higher risk of inferring affect from ambient signal in populations the agent has not been credentialed to model.

Composition With Other Primitives

Empathy outputs feed harm-evaluation directly, but composition with the integrity envelope is mediated. The envelope encodes the agent's own normative commitments and is not modifiable by empathy signal. When empathy-projected harm is high but the action remains within envelope, the harm-minimization subsystem can request alternatives or trigger consent prompts; it cannot unilaterally extend the envelope to forbid the action, nor unilaterally permit a forbidden action because the counterparty seems to want it. Envelope changes proceed only through the deviation-as-mutation procedure, and that procedure does not accept counterparty preference as evidentiary input.

Confidence governance consumes empathy-confidence bands and uses them to modulate the agent's expressed certainty in counterparty-affecting outputs. A recommendation made under high empathy-uncertainty carries explicit hedging, so that a counterparty receiving advice can see the agent's own epistemic posture rather than receiving a falsely confident pronouncement that papers over uncertain inference about how the advice will land.

Lineage binding makes the mechanism auditable. Where a downstream regulator asks why the agent chose a particular formulation, the lineage shows the empathy-derived inference, the bound under which it operated, and the harm-evaluation that consumed it. Where the regulator asks why the agent did not yield to an apparent preference, the lineage shows the bound that prevented the yield. Both directions of inquiry are answerable from the same artifact.

Discovery traversal consumes empathy outputs as ranking inputs but never as admissibility inputs. A traversal may prefer a formulation projected to be better received over one projected to be worse received, but neither formulation's admissibility is determined by the empathy field; admissibility is determined by the integrity envelope and capability envelope alone.

Consent flows are the structural complement to bounded empathy. Where empathy alone is insufficient to authorize an action, the architecture exposes the option of explicit consent solicitation, producing a structured consent record that is itself a first-class lineage event. Empathy informs whether to ask; consent records what was answered.

Prior-Art Distinction

Affective-computing systems generally treat counterparty affect either as a target (behavior cloning, preference imitation) or as an unbounded input that flows into a black-box policy. Both approaches lack the structural separation between sensed counterparty state and agent normative commitment that the cognition architecture establishes. They also lack the influence-bound discipline that prevents empathy from becoming either sycophantic or manipulable, and they typically lack lineage-bound observations, so the auditor cannot inspect what the agent thought the counterparty was feeling at the moment of decision.

Reinforcement-from-human-feedback systems that optimize against counterparty preference signals likewise collapse the distinction between sensing and target. The trained policy effectively encodes the population's expressed preferences as objectives, with no architectural place to point to as "the counterparty's state right now, separately from what the agent intends to do about it." The empathy mechanism preserves that separation as a structural matter, not a training-time abstraction.

The empathy-as-distributed-moral-load mechanism is distinguished by the conjunction of bounded modeling, structural non-overwriting of the integrity envelope, explicit per-class influence bounds, lineage binding, decay and confidence discipline, consent flow integration, and the categorical rejection of counterparty preference as a policy target. The architecture treats empathy as a sense, not a steering wheel, and exposes that distinction in artifacts the auditor can inspect rather than in training-time invariants the auditor must trust.

Disclosure Scope

The cognition patent claims the empathy mechanism as an integrated structure: bounded counterparty modeling with confidence bands, lineage-bound observations, per-class influence bounds, decay and admissibility discipline, structural separation from the integrity envelope, integration with consent flows, and consumption by harm-minimization and negotiation subsystems. Implementations across negotiation agents, clinical companions, mediators, customer-service systems, educational tutors, public-service agents, and adversarial-resilient assistants fall within scope. Licensable embodiments span single-counterparty, multi-counterparty, mediator, adversarial, and public-service configurations, with influence bounds, decay curves, consent defaults, admissibility rules, and per-counterparty memory all configurable through the same policy reference that governs the broader cognitive architecture. Scope extends to any system in which counterparty affect is modeled and the modeled affect is structurally prevented from rewriting the agent's normative commitments.