Weights Learn Experiences, Structures Crystallize Knowledge: Trace-Governed Structural Plasticity for Continual Learning — A Controlled Proof-of-Mechanism Study

Abstract

This record contains the manuscript “Weights Learn Experiences, Structures Crystallize Knowledge: Trace-Governed Structural Plasticity for Continual Learning — A Controlled Proof-of-Mechanism Study”. The paper investigates whether repeated explicit reasoning can alter a learner’s computation graph, not only its numerical parameters. It proposes a trace-governed structural plasticity lifecycle: explicit knowledge and a small core first solve an input and produce a provenance-bearing proof trace; repeated stable traces generate candidate neural circuits whose nodes and edges follow proof dependencies; candidates are shadow-validated before entering the active graph; changes to source rules deactivate dependent circuits and return execution to explicit reasoning. The mechanism is evaluated in a controlled Boolean fact environment across six sequential stages, using five development seeds and five held-out seeds. The learned structural path scores 100% on every seen task at every stage. After four distinct skills, a fixed shared network reaches 80.5% and 88.5% mean accuracy on matched-distribution cases in the two seed groups. A generic growing expert bank also retains 100% on those cases, so structural isolation, not proof traces alone, explains in-distribution retention in this setting. On exhaustive held-out fact combinations, trace-governed circuits retain 100%, while generic growth reaches 76.3% and 77.3%. Supporting tests verify condition-bearing neural topology, rollback after a hidden rule change, and active diagnosis when final-answer feedback is causally ambiguous. The work is a controlled proof-of-mechanism study. It does not claim open-ended language learning, unrestricted task discovery, or a general efficiency advantage. The benchmark is narrow and synthetic, the controls are intentionally simple, and the structural contribution is not fully isolated from richer intermediate supervision.