Maintaining Reusable Learning Maps in AI Ecosystems: Provenance, Requalification, and Reconstruction

Retaining a trained artifact does not necessarily retain the information needed to identify applicable reuse or repair. This paper studies a versioned learning map linked to formation history, applicability, evidence, dependencies, and tested reconstruction routes. Retrieval, adaptation, provenance-based maintenance, justification tracking, and dependency invalidation are established substrates, not new principles claimed here. Five adjudication tests have three roles: M1 and M2 evaluate asset viability through informative retention and the net value of active requalification; M3 tests incremental diagnostic specificity; RC1 and RC2 test reconstruction mechanisms involving repair scope and learned relational history. The narrower hypothesis is that maintained dynamical and cross-level descriptors improve held-out repair selection beyond equally informed case-based and generic predictive alternatives, and that recorded co-learning history helps identify when relations are reusable assets or correction burdens. Diagnosis predicts admissible repairs rather than equating the location of an injected fault with its necessary repair site. Attractor and meta-attractor maps are a dynamical specialization, not labels for every skill or feedback process. An analytical contraction example distinguishes equilibrium identity from timely service. Complete-policy comparisons include mapping, upkeep, failed attempts, and coexistence, with cached lazy requalification as a substantive rival. Predictions, rivals, and measurement conditions remain prospective; no new empirical result is reported. Learning-lineage differentiation and recursive cross-scale transfer remain separate research questions, not premises of this manuscript. Evidence status: conditional theory, operational specification, analytical illustration, and prospective tests. No executed learning experiment or completed empirical preregistration is reported. This manuscript is a scope-separated rewrite of the author's 42-page combined manuscript "Maintaining Reusable Attractor Maps in AI Ecosystems" (Submission Version 1.0, 12 September 2026), which remains unpublished and unchanged; definitions and proposed tests were reorganized for an independent research question, so some measurement and cost conventions are shared with the companion paper. Companion paper: Learning-Lineage Differentiation in AI Ecosystems (https://doi.org/10.5281/zenodo.22722741). The paper derives from the author's broader Information Ecosystem Theory, which also remains unpublished at the time of this release. It is written to stand alone.