The None-of-the-Above Bayesian Hypothesis Controller: Open-World Model Selection with Residual Discovery and Typed Scientific Transitions

Scientific inference systems usually assume that the true explanation is contained somewhere inside a predefined hypothesis set. Under that closed-world assumption, posterior probability must eventually concentrate among the hypotheses supplied to the system even when every available model is wrong. The none-of-the-above residual hypothesis, the probability assigned to mechanisms outside the current named set, has been recognized in Bayesian epistemology for 50 years (Salmon 1975; Earman 1992; Howson & Urbach 2006). Prior treatments identify this catch-all construct as a conceptual requirement of coherent Bayesian confirmation; they do not supply the operational machinery that would make it computationally enforceable.