Affective State for Crisis Response Agents

Regulatory Framework

The Robert T. Stafford Disaster Relief and Emergency Assistance Act establishes the statutory basis for federal disaster response in the United States and triggers FEMA's authorities under presidentially declared emergencies. The National Response Framework operationalizes Stafford by defining response doctrine, Emergency Support Functions, and the relationships among federal, state, tribal, territorial, and local responders. Within NRF, the National Incident Management System and Incident Command System provide the management structure under which all responders, including AI-augmented decision support and communication agents, must operate. NIMS is explicit that command, coordination, and public information functions require trained, predictable behavior from every component participating in the response, and AI agents are not exempt from this expectation.

The European Union Civil Protection Mechanism, governed by Decision 1313/2013/EU and its successors, coordinates response across member states and increasingly with neighboring countries through standardized request, deployment, and information-sharing protocols. The United Nations Disaster Assessment and Coordination system provides similar standardization for international humanitarian response under the UN Office for the Coordination of Humanitarian Affairs. Both frameworks presume that information products entering the coordination network, whether human-generated or AI-generated, conform to doctrinal expectations for tone, urgency calibration, and uncertainty disclosure.

The International Federation of Red Cross and Red Crescent Societies' SPHERE Standards establish minimum humanitarian protection and assistance norms, including the Core Humanitarian Standard's commitments to communication that is timely, relevant, accessible, and respectful of the dignity of affected populations. AI agents communicating with disaster-affected populations are subject to these standards in any operation conducted by SPHERE-aligned organizations, which include nearly all major humanitarian responders.

The U.S. Coast Guard's National Search and Rescue Plan and the international IAMSAR Manual jointly issued by IMO and ICAO govern search and rescue operations. Both presume that decision support tools maintain appropriate operational tempo and that communication artifacts produced during SAR coordination meet the standardized format and tonal expectations that air, surface, and shore SAR units rely upon.

Joint Commission accreditation standards for hospital emergency management, particularly EM standards on incident command and EC standards on emergency operations, presume that decision support and communication tools used in hospital command centers behave predictably under surge conditions. NIOSH first responder mental health guidance, while not regulatory, is increasingly cited in occupational health adjudication and informs reasonable-care expectations for technology that interacts with responders during high-stress operations, including the obligation not to amplify responder distress through inappropriate AI communication tone.

Architectural Requirement

The combined doctrinal and regulatory surface yields a concrete architectural requirement: a crisis response agent must maintain its communicative emotional posture as inspectable, governed state, not as a function of the underlying language model's response to whatever input most recently arrived. NIMS, SPHERE, and IAMSAR all presume that the entity producing communications has an identifiable, predictable disposition; an AI agent whose tone reconstructs from each input cannot be that entity.

The requirement decomposes into four properties. First, the agent's emotional posture must persist across the duration of an incident, with the posture at any moment derivable from the posture at the prior moment plus documented updates from intervening events and elapsed time. Second, the posture must be bounded by governance constraints that prevent pathological configurations, particularly the failure modes that crisis communication doctrine recognizes: panic propagation, false reassurance, urgency fatigue, and inappropriate empathy collapse. Third, posture transitions across incident phases, from acute response through stabilization to recovery, must be explicit and auditable, not implicit in the language model's reaction to changing input distributions. Fourth, the posture must be observable at every moment by incident commanders, who under NIMS hold authority over information products and need to be able to direct the agent's emotional configuration just as they direct any other component of the response.

Prompt engineering and per-message tone modulation cannot meet these requirements. A system prompt is not state; it does not survive a context truncation or a model swap, and it cannot be authoritatively read by the incident commander as the agent's current configuration. Per-message classifiers produce local tone judgments that do not aggregate into a coherent posture. The architectural requirement is for a separate, deterministic affective layer that runs alongside the model and is owned, in the doctrinal sense, by the incident command structure.

Why Procedural Compliance Fails

The dominant approach to AI in crisis response has been procedural: pre-approved templates for public messaging, human review queues for AI-drafted communications, post-incident review of AI behavior, and content-policy guardrails on language model outputs. Each is necessary but structurally insufficient under the operational tempo and stakes that NIMS, SPHERE, and Joint Commission accreditation contemplate.

Pre-approved templates fail because they cannot adapt to the conditions that warrant AI assistance in the first place. The events for which templates are most needed are those least anticipated, and the value of an AI agent in a crisis is its ability to compose appropriate communication for circumstances no template captures. A template-only approach reduces the AI to a retrieval system and cedes the domain in which doctrinal and humanitarian expectations are hardest to satisfy: novel composition under time pressure.

Human review queues fail under operational tempo. SAR cases, mass casualty incidents, and rapidly evolving public safety events do not afford the latency that even efficient human review introduces. Jurisdictions that have attempted to require human review of all AI-generated public alerts have observed that responders begin bypassing the AI altogether rather than accept the latency, which negates the AI's contribution and produces selection effects in which AI is used only when speed does not matter.

Post-incident review fails because it is retrospective. After-action reports identify systemic improvements but do not prevent inappropriate communication during the incident under review. NIMS and Joint Commission accreditation increasingly emphasize real-time governance of decision support tools, not retrospective evaluation, as the standard for acceptable AI integration.

Content-policy guardrails fail because they target outputs, not posture. A guardrail that prohibits certain phrases does not constrain the gradient of urgency drift across a sustained incident, the empathy collapse that occurs when a model is repeatedly fed distressing input, or the inappropriate transition from directive to reassuring tone before the operational situation has actually transitioned. These are posture failures, and posture is precisely what content policies cannot govern.

Procedural overlays surround a sound architecture; they cannot substitute for one. The doctrinal expectation that crisis communication artifacts come from an entity with a governed emotional disposition requires that disposition to exist as an architectural element, not as the emergent average of moderated outputs.

What AQ Primitive Provides

Affective state as a deterministic control primitive provides crisis response agents with persistent emotional fields tuned for emergency operations and bounded by governance configured to crisis communication doctrine. Each agent maintains named fields including urgency, situational stability, trust toward affected populations and toward responders, fatigue tracking for sustained operations, and domain-specific fields such as evacuation pressure for evacuation coordinators or triage tempo for medical command agents. These fields are structural variables maintained by the agent runtime, not generated by the language model, and they are exposed to the model as input on every turn and to the incident commander as observable state at every moment.

Asymmetric update rules reflect the empirical dynamics of operational stress. Urgency rises rapidly in response to incoming severity but decays slowly, producing the operational tempo that responders expect. Trust accumulates slowly through congruent, accurate communication and degrades quickly on observed inaccuracy, mirroring the dynamics that SPHERE Core Humanitarian Standard implicitly assumes. Fatigue accumulates monotonically over sustained operations and only resets through documented rest cycles, providing the structural basis for NIOSH-aligned responder welfare communication.

Valence stabilization is enforced as a governance layer above the field dynamics. Hard bounds prevent the urgency field from producing communication that crosses the line from directive to panicked, regardless of how severe the incoming reports become. Cross-field constraints prevent pathological combinations such as maximum urgency combined with collapsed empathy, which is the configuration most associated with communications that meet NIMS criteria for content but violate SPHERE expectations for dignity. Rate limiters prevent posture transitions that exceed the rate at which incident commanders can supervise them, ensuring that the agent does not silently shift from acute-phase to stabilization-phase tone before the commander has authorized that transition.

Phase transitions across acute response, stabilization, and recovery are themselves first-class state. The agent does not infer the phase from input distributions; the phase is a configured variable, settable by the incident commander, that biases field dynamics and bounds. During acute phase, urgency bounds are wider and stability bounds are tighter, producing directive communication that resists informational noise. During stabilization, the bounds invert, producing more detailed informational communication and greater responsiveness to questions. During recovery, empathy bounds widen and patience-related fields are given longer time constants, producing the sustained, frustration-tolerant communication that recovery operations require.

Multi-agent coordination across response domains uses a configured contagion graph that defines which fields propagate between which agents. The medical command agent's urgency propagates to evacuation coordination but not to public information; the public information agent's perception of population anxiety does not propagate as urgency to infrastructure assessment. The graph is itself configuration, owned by the incident command structure, and modifiable as the response evolves. The result is coordinated emotional posture across the response system that mirrors the doctrinal coordination NIMS requires of the human command structure.

Compliance Mapping

Each doctrinal and regulatory obligation maps to a specific aspect of the affective state primitive. NIMS and ICS expectations of predictable component behavior under command authority map to the observable fields and the incident commander's read and write authority over them. NRF Emergency Support Function communication standards map to the phase-configured bounds and the audit log of posture transitions across the incident timeline.

EU Civil Protection Mechanism and UNDAC interoperability presume that information products conform to standardized tone and uncertainty disclosure norms; the affective configuration enumerating the bounds of urgency and the rules for uncertainty acknowledgment provides the documentation that satisfies these presumptions across multinational coordination.

SPHERE Standards and the Core Humanitarian Standard map to the empathy bounds, the trust dynamics, and the dignity-preserving constraints in the governance layer. The auditable posture log allows humanitarian organizations to demonstrate to donors and to affected populations that AI-mediated communication maintained CHS-aligned tone throughout the operation.

USCG SAR Plan and IAMSAR communication conventions map to the agent's directive-tone bounds during acute phase and to the structured handoff dynamics across SAR phases. Joint Commission EM standards for hospital emergency management map to the surge-mode configuration of the agent and to the incident commander's authority to read and direct the agent's emotional posture during command center operations.

NIOSH first responder mental health guidance maps to the fatigue-tracking fields and the configured constraints on agent communication directed at responders showing fatigue indicators. The agent does not amplify responder distress; it adapts its directive pressure downward and its support upward as fatigue accumulates, and the configuration enumerating these adaptations is available for occupational health audit.

Adoption Pathway

Adoption of affective state as a control primitive in crisis response proceeds in four phases that align with the maturity of emergency management practice in an adopting organization. The first phase is exercise integration: the affective layer is deployed in tabletop and functional exercises, with field values and posture transitions logged but not yet exposed to live operations. This phase produces the empirical baseline needed to calibrate update rules, decay constants, and phase-transition triggers, and it familiarizes incident commanders and public information officers with the observable state interface.

The second phase is supervised live deployment: the agent is used in low-severity incidents under direct incident commander supervision, with the affective state visible on the incident management dashboard and with the commander holding write authority over phase transitions and posture bounds. This phase produces the operational data needed to demonstrate to accrediting and regulatory reviewers that the affective layer behaves as documented under live conditions.

The third phase is full doctrinal integration: the affective configuration is incorporated into incident action plans, the multi-agent contagion graph is documented in the communications annex, and the agent's affective state log is integrated with the incident management system's record of operations. The configuration is brought under the same change management process that governs other doctrinal artifacts, and operator training reflects the structural meaning of fields, bounds, and phases.

The fourth phase is interagency and humanitarian engagement: the affective architecture is presented to FEMA, EU CPM coordinators, UN OCHA, IFRC SPHERE auditors, Joint Commission surveyors, and partner agencies as the substantiation backbone for AI-mediated communication during multi-jurisdiction or international response. At this phase, the agent is no longer a sole-source product of the deploying organization; it is a coordination-ready component whose emotional behavior is documented in terms that partner doctrine recognizes.

The pathway is incremental and reversible at every phase. The deterministic affective layer is independent of the underlying language model and survives model upgrades, providing a stable doctrinal and accreditation surface even as the generative substrate evolves. This stability is itself a property that emergency management leaders value, distinguishing affective state governance from architectures whose emotional behavior is implicit in the model and therefore subject to silent change with every model update.