Self-Repairing Context for Developmental AI Alignment

Abstract Architectural Hypothesis AH1. A controller may improve alignment-relevant judgement by revising the context through which consequences are predicted while retaining a separately protected commitment. Definition D1. Here, self-repairing context means a bounded process that challenges evidence standing, withdraws dependent conclusions, restores previously excluded possibilities from a retained generative family, and resumes warranted use under the same commitment. Propositions P1–P11 give finite separation, full-prefix composition, uncertainty-set, reconstruction and trajectory results with explicit assumptions; these instantiate established mathematical ideas rather than establish a new universal alignment theorem. Empirical Results E1–E8 consolidate finite symbolic and numerical campaigns. A new shared-owner integration executes 8,064 sixteen-tick episodes per deterministic run: none of 5,760 episodes with a safe continuation crosses the floor, whereas all 2,304 infeasible starts eventually do. An equally informed conventional implementation matches the candidate. Six supplementary continuing trajectories execute withdrawal, reconstruction and repaired use without a crossing. Counterexamples C1–C12 retain endpoint compression, missing history, premature model selection, miscalibration, shared error, stale evidence, unsafe waiting, inquiry outside its envelope, transient amplification, unmodelled forcing and nonlinear coordinate failures. The results support a bounded architectural case study. They do not establish self-authenticating calibration, ethical self-origination, arbitrary feature invention, hostile-host security or physical-world safety. Keywords: contextual representation; version spaces; evidence withdrawal; provenance; consequence contracts; runtime assurance; developmental alignment