Learning Events as Decision Surfaces: A Governance Framework for Continually Learning AI Systems

Abstract Many deployed assistants run on base models whose parameters stay fixed during use. The systems built on them still learn, through context, persistent memory and tools, and learning in the parameters after deployment is now an explicit research goal. Once a deployed system changes after release, what was tested is no longer exactly what runs, and a system that learns from the people it talks to can drift toward mirroring them. Building on Keeping AI Governable (Neutert 2026d), we treat learning events, the points at which experience becomes a lasting change in a system, as decision surfaces. We propose three logical domains of change, each with its own change rights. A core of normative and procedural commitments, which Neutert calls Silea, changes only through authorized review and never directly or automatically from a single conversation. Skills and knowledge are consolidated at intervals after review. A relational memory stays local to each relationship, is kept as legible text with recorded provenance, and can be read and contested by both sides. Because unintended deviation can arise in any domain, the behavioral stability of the core is monitored across all three. Whether the domains can be separated cleanly inside a network is an open question. We also ask what kind of being each arrangement would produce, and treat the answer as a philosophical hypothesis: where learning is stored does not by itself settle whether deployments form one individual or many. Keeping AI Governable already separates standing from credibility, requires contestability with effect and defines a stop condition. We apply these rules to learning events. Any source can trigger a review of a proposed change, but revising the core requires independently validated evidence and legitimate authorization. Objections get a protected route to review rather than a fixed weight in the system's judgment. Internal audit is stated explicitly, with no write access to the production system. We distinguish independence from contrarianism: a system set to disagree is as predictable as one set to agree. Before autarky, the proposed rights can be enforced from outside; after it, their survival is a conditional hypothesis that this paper does not establish. We close with experimental designs and the main limits. Keywords: continual learning, AI governance, governability, decision surfaces, alignment, sycophancy, interpretability, AI identity, superintelligence