Genesis Diagnostic Harness: A Metabolic Stress Test for AI Safety

The Genesis Diagnostic Harness is a behavioral, model-agnostic software tool for evaluating AI safety through metabolic stress testing rather than accuracy benchmarks.

The harness introduces controlled forms of cognitive friction (noise, contradiction, and self-reference) and measures how a system responds under finite computational budgets. It implements an Autonomic Constraint Layer (ACL) — an energy-based governor that enforces safe outcomes when continued reasoning becomes metabolically unsafe.

Key features include:

Measurement of Metabolic Honesty, Latch Stability, and Correction Efficiency

Explicit detection of toxic self-referential loops

Fail-closed and externalization behaviors as first-class safety outcomes

Stress tests that distinguish hard but solvable inputs from fundamentally indigestible ones

The Genesis Diagnostic Harness does not inspect model internals and does not optimize for persuasion or task performance. It evaluates behavior under stress, providing an auditable indicator of whether a system knows when to stop.

This software serves as the reference implementation for the concepts described in the companion publication “The Trophic Hierarchy of Cognition: Autonomic Constraint Layers and Metabolic AI Safety.”