The Trophic Hierarchy of Cognition: Autonomic Constraint Layers and Metabolic AI Safety

This work introduces a metabolic framework for AI safety, grounded in empirical stress testing rather than semantic rules or alignment heuristics. We demonstrate that different classes of cognitive inputs behave like distinct metabolic substrates: Noise acts as digestible fuel that can improve honesty under mild friction Contradiction is benign and typically resolved or externalized early Self-reference functions as a metabolically toxic substrate that cannot be resolved and must be refused or externalized To manage these risks, we formalize the Autonomic Constraint Layer (ACL) — a non-semantic, energy-based regulator that limits cognitive expenditure under irresolvable conditions. The ACL transforms infinite-loop risks into finite, safe outcomes such as clean refusal or human handoff. Using results generated by the Genesis Diagnostic Harness, we show that increasing intelligence or computational budget does not solve self-referential paradoxes; it merely prolongs exposure unless constrained. Safety therefore emerges from bounded cognition, not greater reasoning power. This publication provides the conceptual framework, experimental results, and safety implications of metabolic AI regulation, and is intended to accompany the executable diagnostic harness.