From Variety Matching to Stress-Bounded Governance: Extending Ashby's Law under Epistemic Degradation

We revisit the classical law of requisite variety and show that variety matching, though necessary, is insufficient to guarantee stability under epistemic degradation and accumulated stress. We place a noisy, delayed observation channel between disturbance and controller, and prove that an irreducible epistemic deficit survives any amount of controller variety. Integrating the resulting regulation mismatch under finite recovery yields a dissipative stress process whose long run behaviour is governed by a single threshold relating mean overload to dissipation ca pacity (the Stress Bounded Control Theorem). We exhibit Ashby compliant systems that are nonetheless supercritical, and we derive a computable bound on admissible decision latency that holds independently of controller sophistication. We illustrate the framework on instrumented language models, where the spectral sensitivity of the per token Fisher information furnishes an observable surrogate for instantaneous mismatch with a demonstrated prospective relationship to failure (median 27 token lead, 85.6% detection). We propose not a replacement for requisite variety but a second, dynamic safety condition complementing it: safe operation requires that mean overload stay below dissipation capacity. The framework relocates the design question for fast, partially observed, contested systems from how much variety the regulator commands to whether it can keep pace.