Skip to content

Emergency Response

Emergency response represents the highest tier (Tier 3) of AI-assisted nuclear operations. During events like a loss-of-coolant accident (LOCA), the full multi-agent architecture is deployed to support operators managing rapidly evolving, safety-critical conditions.

Full Pattern 9 Deployment

Emergency response deploys the complete Pattern 9 (Shared Room) configuration from multi-agent-patterns, typically with 7 specialised agents:

  • Domain experts: Specialised in thermal-hydraulics, reactor physics, electrical systems, and other relevant disciplines
  • Adversarial agent: Challenges consensus, specifically designed to resist sycophancy and group-think
  • Projection agent: Coupled to simulation codes like relap5 via simulator-coupling to provide physics-based prediction of future plant states
  • Synthesiser: Integrates outputs from domain experts into coherent assessment
  • SA Bridge: Translates multi-agent outputs into operator-comprehensible summaries for situation-awareness maintenance

Mandatory Controls

Emergency tier operations require:

  • Governance gates: All safety-significant recommendations require explicit operator approval before action (see human-authority)
  • Mode escalation: System transitions from normal to abnormal to emergency modes, with each transition activating additional agents and controls
  • Parallel execution: Time-critical emergency scenarios demand parallel agent processing — sequential execution is too slow (see delivery-modes)

Mode Escalation

The system operates in three modes:

  1. Normal: Subset of agents active, standard advisory
  2. Abnormal: Additional agents activated, increased monitoring frequency, enhanced situation-awareness support
  3. Emergency: Full Pattern 9 deployment, all governance gates active, maximum agent complement, parallel execution mandatory

Escalation criteria are defined in advance and linked to plant conditions, not AI assessment — the AI does not decide when to escalate its own authority.

Critical Design Principle

During emergencies, AI value is highest but so is risk. The system must provide genuinely useful analysis without becoming a distraction. Operator attention is the scarcest resource during an emergency, and every piece of AI output that requires operator evaluation consumes that resource. The flow gate mechanism ensures operators receive information at a manageable rate even when agents produce output rapidly.