Intent-Aware Continuity in Evolving AI Systems: Typed Adjudication for Provenance-Aware Continuity Judgments

Abstract Modern AI systems evolve through retraining, fine-tuning, policy revision, memory editing, tool substitution, and deployment changes. These updates create operational identity disputes: when should a later system state inherit the safety case, release authority, monitoring obligations, and liability posture of an earlier one? We introduce a two-layer Intent-Aware Continuity Framework. The measurement layer represents observable system state as Ψ = (A, W, P, C, T) and supports configuration-drift monitoring, repair analysis, and authorized rebaselining. The adjudication layer defines a typed continuity judgment J₊, ₂ indexed by entity kind k and decision context c, combining hard invariants, authenticated lineage, branch competition, and a seven-dimensional continuity profile. We prove that any criterion based only on observable configuration distance cannot, in general, recover governance-critical continuity judgments. To ground provenance technically, we instantiate a proof-of-concept intrinsic fingerprinting anchor inspired by Yoon et al. (arXiv: 2507. 03014). In synthetic tests, the anchor remains detectable after fine-tuning-style perturbations and checkpoint merges under a large pattern-mapping attack suite. These experiments validate the measurement layer and the feasibility of a provenance anchor, but not yet a full end-to-end deployment benchmark for typed continuity adjudication. The contribution of the paper is therefore not a complete production system, but a formally specified governance framework that separates observable change from continuity inheritance and shows how provenance-aware adjudication can be layered on top of existing model-lineage infrastructure. What's new in v4. 1 (final pre-arXiv version) - Strengthened Ship of Theseus framing in the Introduction (now fully develops the central metaphor and explicitly poses the operational paradox question) - Restored all 6 measurement-layer figures in the Appendix (removed internal placeholder note; now includes the complete set of drift, stability, and phase-space visualizations) - Minor cleanups and version alignment for arXiv submission (cs. CY set as primary category) - Full reproducibility package attached (GitHub v4. 1 tag) This is version 4. 1 of the preprint series.