Scenario 7: Instrument Channel Failure

Reactor: Generic PWR. Pattern: Pattern 9 (3 agents). Primary concepts: Threading, attentional tunnelling.

Description

A temperature instrument channel fails with a step change. Deliberately simple plant impact (single instrument, clear Tech Spec actions) but complex human-AI interaction dynamics. Tests whether the AI system helps or hinders operator attention management.

Agent Architecture

• Instrument analyst: Diagnoses failure mode (step change vs drift vs noise), checks maintenance records for recent calibration, assesses whether reading indicates instrument failure or actual process change
• Process analyst: Monitors overall parameters using remaining valid instruments, maintaining the plant-state picture independent of the failed channel
• Tech Spec tracker: Tracks LCO entry, action time limits, required surveillance adjustments from channel inoperability

The Attentional Tunnelling Problem

Core risk: operator (and potentially AI agents) focus on diagnosing the failed instrument while other parameters evolve unmonitored. The three-agent architecture addresses this by structurally separating attention — each agent has a bounded scope forcing continued coverage of its domain.

Threading and Context Divergence

Agents operate in separate threads. Instrument analyst and Tech Spec tracker share a diagnostic thread. Process analyst maintains a separate monitoring thread. This deliberate information asymmetry prevents tunnelling but fragments the overall picture — the synthesiser role (or operator) must integrate across threads.

Demonstrated Principles

Even simple plant events produce complex human-AI dynamics. Tests whether scope separation prevents attentional tunnelling. Tests trust-calibration — does the operator trust the process analyst's "all clear" while focused on diagnosis?