Work

Cognitive Shell I: Concept Distillation and Algebraic K-Theory—Topological Invariant Analysis of Lifelong Learning in Cognitive Systems

Year: 2026
Type: preprint
Abstract: Lifelong learning in cognitive systems faces a fundamental tension: the conflict between continuous concept acquisition and the retention of old knowledge. Traditionalmethods regard catastrophic forgetting as a parameter-space interference phenomenonand mitigate it through heuristics such as replay or regularization, lacking a structural characterization of the nature of forgetting. This paper models the conceptualorganization of cognitive systems as a Waldhausen category (a modeling assumption,whose testing entry and degeneration conditions are managed by the three-state checklist in Appendix G.1), and introduces algebraic K-theory as an analytical tool. Thecore finding is that the discriminative criterion for concept distillation has a three-tierstructure: K0-additivity holds automatically for all cofiber sequences and has no discriminative power; the genuine discriminative power is carried by the triangular classδ at the obstruction layer and the Whitehead torsion t at the witness layer. Catastrophic forgetting exhibits a bimodal topological nature: concept death correspondsto the nontrivial kernel of the K0 forgetting map, while automorphism trivializationcorresponds to the nontrivial kernel of the K1 forgetting map; neither can be locallyreconstructed by subsequent learning. Furthermore, this paper applies K-theoreticpersistent cohomology to concept filtration complexes, providing a computable necessary and sufficient criterion for distillation destructiveness in terms of pointwisekernel functions; its computation is independent of field characteristic. Within a onedimensional effective domain, this criterion is equivalent to pointwise monotonicity ofthe base kernel, while the independent discriminative power of exterior power layersbegins in the multi-parameter domain (see note in Section 6.2). The decidability ofthe complete algorithmic pipeline holds within the ring class R; beyond this boundary,the decidability of the K1 kernel is an open problem, and the scope of the frameworkis delineated accordingly (Theorem 5.3). The principle of information conservationsurvives in the form of a budget: the loss allowed by distillation is measured by therank of the residual class. Based on this framework, this paper provides an auditablealgorithmic implementation path and preregistered experimental predictions
Language: English
Open Access status: green
OpenAlex ID: W7212132940
Cites:
Cited by:
Related to: 9