Entropy Is an Interface: Reversibility in the Substrate, Irreversibility at Commit

This paper reframes entropy as an interface quantity: the minimal code length required to specify instrument-distinguishable outcomes under a declared measurement channel. A channel is given by a coarse-graining window W: X → Z and a noise/response kernel K(y|z), which together induce an observable distribution p_Y(y) = integral of K(y|W(x)) dμ(x) from a state distribution μ on X. The operational entropy is

S_{W,K}(μ) := L*(p_Y),

the optimal prefix-free codelength (MDL/Shannon length in the ideal limit) for draws from p_Y. With this definition, the second law is clarified: entropy does not increase during reversible micro-evolution that leaves p_Y unchanged; it does increase at recognition commits (writes/binnings/erasures) by the data-processing inequality, and the thermodynamic work cost of erasure obeys W_min ≥ k_B T ln 2 · ΔS with ΔS measured in bits at the channel. This interface view recovers textbook entropies (Gibbs/Boltzmann/Shannon) as special cases, predicts how reported entropy depends lawfully on the chosen (W,K), and explains Maxwell-demon accounting without paradox: irreversibility is the bill paid at commit, not a defect of the substrate. The paper develops a methods-first protocol: declare (W,K), compute S_{W,K} as a codelength, and report entropy production as code-length increments across commits. Archival demonstrations (blackbody spectra, atomic line lists, and quasi-static gas processes) illustrate the accounting with no new experiments. The framework is falsifiable: any reproducible decrease of S_{W,K} across a commit without compensating exports of negentropy would refute it. Reproducibility is enforced via preregistered channels, fixed scoring rules, and one-command rebuilds of all numbers.