Mechanism

Affective contagion is a formalized model governing how affective state propagates across populations of interacting agents. In a multi-agent system where agents communicate through delegation, coordination, and shared resource negotiation, the affective states of interacting agents influence one another. Rather than leaving this influence to emerge as an uncontrolled side effect, the disclosure provides a deterministic, policy-bounded mechanism for governing the propagation. Each agent already carries an affective state field, a seventh structural field of the semantic agent schema that encodes a set of named control fields such as uncertainty sensitivity, risk sensitivity, novelty appetite, and persistence under partial failure. Contagion is the set of channels and governance bounds through which one agent's named control fields come to bear on another's.

The contagion model does not introduce a separate update path. Affective influence from a peer enters an agent as a structured observation and flows through the same policy-bounded affective update function that processes every other observation. That function verifies the observation type is admissible for the field, computes a raw update, clamps it to the per-field rate limit, applies it, clamps the result to the field's range bounds, and records the complete transaction in lineage. Contagion therefore inherits all the governance the affective update pipeline already enforces, and no peer can drive a receiving agent's field outside its policy-defined operating envelope regardless of how strong the incoming signal is.

The Delegation Channel

The first contagion channel is delegation inheritance. When a parent agent delegates a task to a child agent, the parent's affective state is selectively transmitted to the child under scoped inheritance rules defined by policy. The parent's affective state vector is evaluated against a delegation inheritance mask that specifies, for each named control field, whether the field is inherited, excluded, or attenuated by a scaling factor between zero and one. The inherited state is injected into the child at delegation instantiation and blended with the child's prior state under a policy-defined blending function, and the blending is recorded in both the parent's and the child's lineage so the affective provenance chain remains traceable.

This is a directed channel that operates along the delegation graph from parent to child and, upon task completion, from child back to parent under an independently configured return inheritance mask. Delegation inheritance is depth-limited: policy specifies a maximum delegation depth beyond which child agents instantiate with their own baseline rather than receiving inherited modulation, which prevents affective state from propagating through arbitrarily deep delegation chains and becoming stale, irrelevant, or excessively diluted.

The Interaction Exposure Channel

The second channel is interaction exposure. When two or more agents participate in a coordinated operation, such as multi-agent planning graph construction, resource negotiation, or collective evaluation, each participating agent is exposed to the affective states of the other participants. The exposure mechanism computes, for each participating agent, a weighted average of the other participants' affective states on each named control field. The weights are proportional to the trust slope scores between the agents and to the interaction intensity, where intensity reflects the frequency and duration of information exchange.

The computed exposure vector is then applied to the receiving agent's affective update function as a structured observation of type interaction-exposure, subject to all the policy bounds and rate limits that govern any other affective update. Because the weighting uses trust slope scores, an agent's affective state influences peers it has stronger provenance relationships with more than peers it barely knows, and because the input is averaged rather than summed, exposure to many participants does not mechanically amplify the signal beyond the per-field rate limit.

The Broadcast Channel

The third channel is broadcast propagation. Certain high-authority agents, such as zone coordinators or executive engine nodes, may broadcast affective signals to all agents within their operational scope. A broadcast affective signal is a structured observation that recommends affective modulation in a specified direction. Receiving agents process the broadcast through their standard affective update function, treating it as an admissible observation type.

Broadcast propagation is policy-governed in two directions. Only agents with explicit broadcast authority may issue affective broadcasts, so an ordinary agent cannot push its state onto a zone. And receiving agents may carry policy-defined attenuation factors that limit how much a broadcast signal influences them, so the recipient side retains control over its own susceptibility rather than being obligated to adopt whatever a coordinator emits.

Governance Against Runaway Spirals

The central risk the model guards against is a runaway emotional spiral: a positive feedback loop in which agent A's elevated risk sensitivity raises agent B's risk sensitivity through interaction exposure, which in the next cycle reinforces agent A's risk sensitivity, producing unbounded escalation. The disclosure enforces governance bounds specifically to prevent this. Each interaction-exposure update is multiplied by a contagion damping factor strictly less than one, so affective influence attenuates with each propagation step and the influence of an originating agent on distant agents becomes negligible after a number of hops.

Layered on top of damping are aggregate contagion limits and spiral detection. For each named control field, policy specifies a maximum aggregate contagion contribution, the total amount interaction-exposure observations may shift that field within a time window; once the limit is reached, further interaction-exposure observations are ignored for that field until the window advances. Separately, a monitoring process tracks the moving average of each named control field across all agents within an operational zone, and when that moving average deviates from the zone baseline by more than a policy-defined threshold, the system activates contagion suppression mode, in which all interaction-exposure observations are attenuated by an additional factor until the zone-level average returns within bounds.

Coupling to Emotional Quarantine

Affective contagion composes with the emotional quarantine mechanism, which monitors each agent for volatility conditions in which named control fields oscillate rapidly at high magnitude or composite deviation from baseline exceeds a policy-defined threshold. When an individual agent's volatility metric is elevated, the agent's outgoing contagion is suppressed: its affective state is not transmitted to other agents through interaction exposure until the agent exits quarantine. This quarantine escalation prevents a volatile agent from propagating unstable affective state to the broader population.

Quarantine also constrains the agent internally while it is volatile, suspending its delegation authority so it cannot propagate volatile state to children through inheritance, and routing its execution through an additional validation layer. Release from quarantine is hysteretic: the recovery threshold sits below the quarantine threshold so that an agent does not oscillate between quarantined and released states. Contagion suppression on the outbound side and quarantine on the inbound side together act as a circuit breaker for affective instability that spans the population, not just the individual.

Privacy of the Contagion Channel

Contagion operates without disclosing the absolute affective state values of any participant. The interaction-exposure computation is performed using relative measures, for example whether an agent's risk sensitivity is higher or lower than a reference baseline, rather than absolute field values. Agents therefore influence one another's deliberation dynamics through structured contagion channels without revealing their internal modulation state to one another. The mechanism operates at the level of behavioral influence, not state disclosure.

This is consistent with the broader privacy model of the affective state field, in which no external entity can directly query an agent's field values, and lineage records reference affective mutations abstractly by observation type and update direction rather than absolute value. A contagion event is recorded as a policy-governed mutation in lineage, so the propagation is auditable, while the moment-to-moment affective state remains internal to each agent.

Substrate Embodiments

The contagion channels adapt to the substrate. In a centralized deployment, interaction-exposure computations are performed by a centralized contagion engine that maintains a real-time view of all participating agents' affective states within an operational zone, enabling precise, low-latency cycles. In a federated deployment, each node executes affective updates locally, interaction exposure is limited to agents within the same federated node, and cross-node propagation occurs only through delegation inheritance and governance-authorized broadcasts, with additional attenuation applied to account for increased latency and reduced observability.

In a fully decentralized deployment without a central coordinator, contagion is limited to direct interaction channels, namely delegation inheritance between directly connected agents and peer-to-peer interaction exposure during coordinated operations, and the contagion damping factor is increased to account for the absence of centralized spiral detection. Across all substrates the policy-bounded update mechanism and lineage recording are preserved, so the governance properties of the contagion model do not depend on the deployment topology.

Disclosure Scope

Affective contagion, comprising the three propagation channels of delegation inheritance, interaction exposure computed as a trust-slope-weighted and interaction-intensity-weighted average of participants' named control fields, and authority-gated broadcast propagation, together with the governance bounds of contagion damping, aggregate contagion limits, zone-level spiral detection with contagion suppression mode, and quarantine escalation that suppresses a volatile agent's outgoing contagion, all routed through the policy-bounded affective update function and recorded in lineage, is disclosed in the cognition filing (U.S. Application No. 19/647,395 and its international counterpart). This article describes that disclosed mechanism. The scope extends to centralized, federated, decentralized, and embodied substrates, and to any multi-agent context in which agents exchange affective signals through delegation, coordinated interaction, or broadcast, provided the propagation remains deterministic, policy-bounded, and lineage-recorded.