Typed Repair: EMNIST from $\lambda$-Definable Evaluation and Specialization Gate (Draft)

This repository contains a preliminary working draft of a paper developing a finitary, semantics-locked learning and evaluation artifact for EMNIST digit classification, together with a decidable specialization (equivalence) gate for residual evaluators. The manuscript is intended as an exploratory contribution toward making learning pipelines auditable by construction: the central objects (preprocessing lock, witness table, margin-repair loop, cache invariants, and finite-gate equivalence checks) are defined so that core verification predicates are total and decidable on an explicitly finite artifact. Status: Draft / Work in Progress.This release should be treated as a research draft. Several components require further investigation and consolidation, including (but not limited to) formal completeness of the metatheoretic claims, robustness of the semantics-lock assumptions across environments, the generality of the intermediate representation families considered, and the extent to which the certification methodology scales beyond the demonstrated finitary setting. Results reported here should therefore be interpreted as provisional and subject to revision. Feedback, corrections, and pointers to related work are welcome.Contact: www.milanrosko.com