Learning-Lineage Differentiation in AI Ecosystems: A Conflict-Conditioned Theory of Preservation and Adaptation

When an AI service must preserve still-required knowledge while adapting current behavior, role assignment alone does not determine how learning histories should differ. This paper studies the value of separately maintained, role-conditioned learned states under common information, resources and service obligations. L1 predicts a practical-margin crossing in selected differentiated-minus-shared performance across independently calibrated reference-conflict conditions. L2 tests two operative training-coordinate contributions. L3a calibrates useful exchange, while L3b tests differentiated updating under matched delivered evidence, distinguishing future-update effects from accumulated-history effects. A finite-update geometric example derives a conditional benefit despite an exactly representable shared solution and identifies remedies that remove it; an attainment bound states when its ideal margin survives finite learning. Contextual prediction and versioned routing workloads retain shared, hybrid, same-regime and archive-supported rivals. Development search, failed selection, transition work and two distinct timing choices - evidence capture and learning separation - remain explicit. External ablations and comparator, cost and metric contrasts constrain candidate explanations. The contribution is a conditional explanation and discriminating test program for learning organization. The paper reports no trained-system experiment, completed preregistration or established crossover. Note on Version 2.0: this version revises the registered v1.0 (about 9,600 to 13,400 words). It adds a finite-update mechanism with a shared solution and an attainment bound, a versioned routing service and a two-coordinate test recipe, a test of learning that persists after exchange, a secondary test of the shared-remedy explanation, and a section on external evidence, numerical gaps and testable bridges. No empirical result is claimed: the supporting archive records analytic design, inference, math, strengthening and external-bridge checks with their verification scripts, a consolidated review note, an edit log, a source lock, a validation record and SHA-256 checksums. Files: the v2.0 manuscript and that supporting archive. This manuscript is a scope-separated rewrite of the author's combined manuscript "Maintaining Reusable Attractor Maps in AI Ecosystems", which remains unpublished; some measurement and cost conventions are therefore shared with the companion paper. Companion paper: Maintaining Reusable Learning Maps in AI Ecosystems (https://doi.org/10.5281/zenodo.22722942). The paper derives from the author's broader Information Ecosystem Theory, which also remains unpublished at the time of this release.