Entry Loss under Quotient Optimisation: Capacity-dependent Contraction and Underdevelopment of Counterfactual Representations
Version 1.1 (17 September 2026): reproducibility packaging update. This version adds a top-level reproducibility entry point and explicit resource documentation. The five v1.0 evidence archives are preserved byte-for-byte. The CIFAR-10 route keeps its executable source and is now labelled resource-conditional. No scientific result, positive finding or null is changed. Optimisation through a lossy proxy can preserve present task performance while changing which distinctions remain easy to express or learn later. This paper calls that effect entry loss and separates two regimes. In a whitened linear quotient bottleneck, gradient descent provably contracts the directions that a shortcut-solvable objective does not require. Under representational scarcity, shortcut-solvable quotient optimisation can actively contract stable future-task distinguishability. Under adequate width, the same commitment can preserve absolute distinguishability while still diverting development relative to matched alternatives. The distinction matters because the two regimes support different claims. Scarcity produces destruction of a capability; adequacy produces diversion of development without destruction. A confirmatory CIFAR-10 assay at adequate width returns the weaker of the two, and that null is reported as a boundary on the claim rather than omitted. Claim boundary. Public deposition of the authoritative seed-level result bundles is complete. External replication and independent re-verification have not yet been performed, so the numerical claims should be treated as project-canonical rather than independently audited from the manuscript alone. This deposit contains the manuscript and supplement in PDF and DOCX form together with the complete evidence archive: the theorem and assay bundle, distribution-shift replication, natural-data replication, CIFAR-10 confirmatory bundle, seed-level outputs, analysis scripts and checksums.
Authors
- Joel Pearcey
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-09-17
- DOI
- https://doi.org/10.5281/zenodo.22811038
- Primary Topic
- Meta-analysis and systematic reviews
- Type
- preprint