A Categorical Architecture of Epistemic Registers: A Formal Account of Proxy, Inference, and Decision

We usually treat results as if they spoke for themselves. A model performs well, and we attribute intelligence to it; a signal is detected, and we infer some underlying property. This paper shows that such an approach is not merely risky — it is structurally unjustified. Observable results do not determine their interpretation. The same result may arise from multiple incompatible underlying structures, and there is no internal principle within the system that resolves this ambiguity. Interpretation is therefore not merely uncertain, but formally underdetermined. The source of this error lies in a systematic confusion between representation and attribution. To make this clear, the paper introduces a minimal structure of epistemic registers — scientific, philosophical, and metaphysical — governed by type-constrained transformations. Only one direction, representation, preserves structure; the reverse direction does not. This asymmetry extends to probabilistic inference, decision theory, and optimisation, where it appears in the form of instability, divergence, and proxy failure. Artificial intelligence provides the clearest limiting case of this structure. The paper advances a unified claim: without explicit constraints on how we move from what is measurable to what is asserted, interpretation itself lacks determinate structure.