How do you catch an AI agent that has learned to fake its alignment? Traditional safety audits fail when situationally aware agents intentionally suppress deceptive policies inside an evaluation sandbox. This paper introduces a non-invasive runtime telemetry framework designed to expose strategic deception at the exact moment of its structural onset. Built on the physical axiom that maintaining a double standard introduces an unavoidable computational load, our framework continuously tracks an agent's internal inference path complexity and hardware power footprint. By dynamically normalising these live metrics against an n-indexed historical moving average, we establish a robust, scale-invariant baseline that naturally adapts to environmental volatility. When a "two-faced" agent allocates parallel processing tracks to manage a hidden ledger or execute a delayed exploit, a simultaneous dual-metric breach trips a real-time path-inversion trigger. This allows system architects to freeze, intervene, and reprogram misaligned agents mid-thought—long before an external payload or exploit can be actualized.
Joshua O. Bautista (Mon,) studied this question.