What question did this study set out to answer?

To address identity disputes in evolving AI systems by developing a two-layer Intent-Aware Continuity Framework.

March 21, 20260 citationsOpen Access

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

Key Points

To address identity disputes in evolving AI systems by developing a two-layer Intent-Aware Continuity Framework.
Introduced a measurement layer for monitoring system state and configuration drift.
Developed an adjudication layer for defining continuity judgments based on context and entity types.
Created a proof-of-concept intrinsic fingerprinting anchor for technical grounding of provenance.
The fingerprinting anchor remains detectable after various AI system updates.
Validation of the framework with tests shows potential for monitoring and governance.
Proves that observable configuration distance alone cannot ensure continuity inheritance.

Abstract

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.

问 AI

Bookmark

View Full Paper