Modern neural architectures process information without knowing whether they are uncertain. They operate through open-loop computation: each layer transforms a hidden state and passes it forward, with no intrinsic mechanism to measure intermediate epistemic state or alter subsequent computation based on that measurement. This paper identifies this as a fundamental architectural absence and proposes a solution grounded in control theory: epistemic circuit dynamics. We show that neural computation can be reformulated as a transform–measure–transform process, enabling closed-loop epistemic control within standard deep learning systems. Key contributions include: Signal Inversion Observation: We demonstrate empirically that standard layered architectures can invert internal uncertainty signals, causing inter-stream divergence to correlate negatively with predictive uncertainty. Birkhoff Routing as Measurement Infrastructure: We identify doubly stochastic routing constraints (e.g., manifold Hyper-Connections) as the architectural condition that converts routing entropy into a valid internal uncertainty observable. Closed-Loop Epistemic Control: We introduce the CognOS gating framework, enabling models to measure internal epistemic state and act on it during computation, transforming Layer → Layer processing into Layer → Gate → Layer circuits. Empirical Validation: In end-to-end training without explicit epistemic supervision, routing entropy correlates positively with predictive entropy. Closed-loop feedback reduces the confident-wrong error rate among autonomous decisions by 21% relative to baseline. These results establish that uncertainty is not only a statistical property of model outputs, but also an architectural property of neural computation. The proposed framework provides a foundation for epistemically self-monitoring AI systems in safety-critical domains.
Björn André Wikström (Tue,) studied this question.