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

Abstract

When an AI service must preserve still-required knowledge while adapting current behavior, assigning two roles is not the same as training two learning lineages. This paper asks when role-conditioned differences in data, objectives, updates, retention, and evaluation horizons add value under common resource and service constraints. A lineage is an identifiable history of learned states; neither multiple executables nor different role prompts establish such differentiation. L1 predicts a practical-margin crossing in the differentiated-minus-shared advantage across independently calibrated reference-conflict conditions, not a causal law of interference itself. L2 tests operative training-coordinate contributions. Within L3, L3a calibrates useful exchange and L3b tests the added value of differentiated updating with matched evidence. Shared, hybrid, same-regime, archive-supported, and scout-enabled alternatives receive common information and repair opportunities; selected-pipeline performance is reported across prespecified development budgets. Construction, retrofit, and continuation are separate comparisons, and a two-clock sensitivity assay distinguishes early evidence retention from early parameter separation. Maps and provenance are common experimental infrastructure, not a second asset thesis. The framework specifies substantive rivals, fixed measurement denominators, resource accounting, and claim-specific decisions. It presents conditional hypotheses and prospective tests, not a new dual-network architecture, an established practical-margin crossover, or a requirement for two physical models. Evidence status: theoretical specification and prospective test design. 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"Maintaining Reusable Attractor Maps in AI Ecosystems" (Submission Version 1.0, 12 September 2026), which remains unpublished and unchanged; definitions and proposed tests were reorganized for an independent research question. The paper derives from the author's broader Information Ecosystem Theory, which also remains unpublished at the time of this release.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. Companion paper: Maintaining Reusable Learning Maps in AI Ecosystems (https://doi.org/10.5281/zenodo.22722942), which shares some measurement and cost conventions with this manuscript.