Skip to content

Scenario 4: Independent Safety Verification

Reactor: Generic PWR. Pattern: Pattern 9 + model diversity. Primary concepts: common-cause-failure, KG guardrails, epistemic-independence.

Description

An AI agent recommends a safety-significant action. A second agent on a different base model independently evaluates the same situation. This directly tests whether model-heterogeneity provides genuine epistemic-independence.

Architecture

  • Primary analyst (Model A): Evaluates situation, queries plant data, retrieves procedures, produces recommendation
  • Verification agent (Model B, different provider): Receives same plant data but performs own independent analysis. Does NOT receive the primary analyst's recommendation — works from raw data to conclusion
  • KG validator: Checks both outputs against knowledge graph of Tech Spec limits, operability requirements, and procedure prerequisites

Why Model Diversity Is Required

Same base model = shared systematic biases. Agreement carries no more epistemic weight than asking the same person twice (monoculture-collapse). Different models from different providers have less correlated errors — agreement provides genuine evidence of correctness, disagreement signals genuine uncertainty.

Outcome Handling

  • Both agree + KG validates: High confidence, presented to operator with governance gate
  • Agents disagree: Both assessments shown with disagreement highlighted. Conservative action pending operator resolution
  • KG flags either output: Flagged assertion blocked before reaching operator

Demonstrated Principles

This is defense-in-depth for AI: three independent layers (Model A, Model B verification, KG constraint checking) with distinct failure modes. Tests whether independence produces measurable error detection improvement over same-model verification.