Emotional Quarantine and Volatility Management

Mechanism

The mechanism rests on three coupled subsystems: a volatility estimator, a containment gate, and a hysteretic exit controller. The volatility estimator maintains a rolling window of per-dimension affective signals, including valence, arousal, confidence, frustration, and caution. For each dimension it computes an oscillation rate, a magnitude range, and a directional dispersion statistic. These are combined into a composite volatility metric using a policy-defined weighting vector. The metric is recomputed on every cognitive tick or, in event-driven configurations, on every state mutation.

When the composite metric crosses the activation threshold, the containment gate engages. Engagement is atomic with respect to the agent's outbound interface: any pending delegations are halted, in-flight proposals are flagged for elevated review, and the agent's policy reference is rewritten to substitute the quarantine policy for the nominal policy. The substitution is performed at the policy boundary rather than inside individual decision sites, ensuring that no execution path can bypass the constraint by predicating on stale parameters.

Exit is governed by a hysteretic controller. The volatility metric must fall below a lower release threshold, separated from the activation threshold by a policy-defined margin, and must remain below that threshold for a continuous stability dwell time. Only after both conditions are satisfied does the gate release. This prevents the chattering behavior that arises when a single threshold is used and the metric oscillates near the boundary.

Critically, the mechanism composes with promotion-containment. Promotion-containment governs whether an agent's local belief or proposal is allowed to be promoted into a shared or higher-authority context. Emotional quarantine modifies the inputs to promotion-containment by raising the required confidence floor and by injecting the volatility metric itself as a gating predicate. The two mechanisms together enforce a layered invariant: an unstable agent cannot promote, and a stable promotion cannot originate from an unstable evaluator.

Operating Parameters

The activation threshold is expressed in normalized volatility units. Typical deployments place it between 0.55 and 0.75 of the empirical maximum observed during nominal operation, with conservative deployments using lower values to favor early containment. The release threshold is offset below activation by a hysteresis margin of 0.10 to 0.20 normalized units. The stability dwell time ranges from a few cognitive ticks for high-frequency agents to several seconds for slower deliberative agents.

The rolling window length determines the responsiveness-stability tradeoff. Short windows make the metric responsive to genuine spikes but sensitive to transient noise; long windows smooth noise but delay containment. Practical configurations use a primary window of moderate length combined with a short auxiliary window that can trigger containment on its own when extreme magnitudes are observed, even if the primary window has not yet responded.

Quarantine policy parameters specify the elevated promotion threshold, the set of suspended delegation classes, and the additional validation predicates that must be satisfied before any mutation is admitted. These parameters are addressable by reference rather than embedded inline, so that operators can tune containment severity without modifying the agent.

Alternative Embodiments

In a per-domain embodiment, the agent maintains separate volatility estimators and separate containment gates for each cognitive domain it operates in. A volatility spike in the financial-decision domain triggers quarantine for financial actions while leaving navigation, communication, and routine self-maintenance unaffected. This embodiment is preferred where the cost of full quarantine is high and the affective disturbance is plausibly localized.

In a graduated embodiment, the gate is not binary but discrete-multilevel. Crossing the first threshold elevates promotion thresholds; crossing the second suspends a subset of delegations; crossing the third suspends all delegations and engages full external review. Each level corresponds to a distinct policy reference, allowing the response to scale with the severity of the disruption.

In a federated embodiment, the volatility metric is shared with peer agents and with a supervisory orchestrator. Peer agents can adjust their own acceptance criteria for proposals originating from a quarantined source, providing a second layer of containment beyond the agent's own gate. The orchestrator can additionally route work away from the quarantined agent until release.

In a therapeutic embodiment intended for companion and counseling applications, the quarantine state is partially observable to the human counterpart through a calibrated affective signal, and the agent's outputs are routed through a stabilizing transformation that removes high-volatility lexical and prosodic features without altering semantic content.

Composition

Emotional quarantine is most powerful when composed with promotion-containment and with the affective-state observability surface. Promotion-containment provides the upward gate; quarantine provides the conditions under which that gate is tightened. The observability surface ensures that the volatility metric, the gate state, and the active policy reference are all auditable, enabling external systems to reason about whether a given decision was produced under nominal or quarantined conditions.

Composition with delegation governance is also direct: when an agent is quarantined, its delegation tokens are revoked or downgraded, and any peer that holds a delegation from the quarantined agent receives a containment notification. This prevents the disruption from propagating through the delegation graph while preserving the agent's identity and its non-delegated capabilities.

Prior-Art Distinction

Conventional safety approaches treat unstable agents through outright shutdown, rate limiting, or human-in-the-loop escalation. Shutdown is destructive: it discards in-flight cognitive work and forces a cold restart. Rate limiting reduces throughput but does not address the underlying unreliability of decisions made under volatility. Human escalation is unscalable and introduces latency that defeats real-time deployments.

Emotional quarantine differs by preserving the agent's cognitive activity while structurally constraining its outbound authority. It treats affective volatility as a first-class signal with its own measurement, threshold, and exit conditions, rather than as a derivative of confidence or error. And it integrates with the agent's policy reference at the boundary, rather than as a wrapper around individual decision calls, which closes bypass paths that wrapper-style approaches cannot close.

Implementation Considerations

Implementations should preserve the agent's cognitive continuity across quarantine entry and exit. Discarding in-flight plans on entry produces a costly cold-start on release; preserving them allows the agent to resume coherent operation as soon as the gate releases. The plans themselves are not executed during quarantine, but they remain available for reevaluation against the restored nominal policy.

Auditing requirements imply that gate state transitions, the volatility metric trace that drove them, and the active policy reference at the time of every outbound action be persisted to an append-only log. Downstream systems that consume the agent's outputs can then reason about whether a given output was produced under nominal or quarantined conditions, and they can apply additional review to outputs produced near a transition boundary.

Adversarial robustness is a deliberate design goal. An adversary that can manipulate inputs to push an agent into volatility could, in principle, weaponize quarantine to suppress the agent. The mechanism resists this in two ways: by preserving forecasting and planning during quarantine so that the agent can continue to reason about and report on its own state, and by exposing the volatility metric externally so that suspicious patterns of induced volatility can be detected by orchestrators or peers.

Calibration of activation thresholds is performed against an empirical distribution of the volatility metric collected during a controlled bring-up period. The bring-up establishes the nominal range of the metric for the deployed agent in its target environment, and the activation threshold is placed at a quantile of that distribution chosen to balance false-positive containment against missed detections. Recalibration is performed periodically and after any material change to the agent's policy, sensors, or environment, recognizing that the volatility metric is environment-dependent and that a threshold tuned for one operating context will not generalize to another without revalidation.

Disclosure Scope

The disclosure covers any system in which a measured affective or volatility signal triggers a structural containment of an agent's outbound authority through substitution of a quarantine policy reference, with hysteretic exit conditions and composition with promotion-containment. The disclosure is not limited to the specific dimensions of affect described, nor to the specific weighting or windowing strategies used to compute the composite metric. It extends to per-domain, graduated, federated, and therapeutic embodiments and to any embodiment in which the gate is engaged at the policy boundary rather than at individual decision sites. It further covers compositions with delegation governance, with observability surfaces that expose the gate state to peer agents and orchestrators, and with auditing infrastructure that records gate transitions and policy substitutions for downstream reasoning about output provenance.

As described in U.S. provisional application 64/049,409, the quarantine gate is implemented as a structural policy substitution rather than a runtime filter, with the substitution event itself recorded as a signed observation whose lineage composes with downstream agent decisions. This treatment ensures that any output produced while the agent was under quarantine carries verifiable provenance back to the governing policy reference in force at decision time, supporting after-action review, regulatory examination, and the reproduction of constrained behavior in offline simulation without reliance on volatile in-memory state.