Memomind: Criticality-Stratified Learning for Efficient Computation Orchestration

Criticality-stratified learning targets strict accuracy on a critical subset of inputs while allowing relaxed accuracy on others, matching real systems that route heavy compute only when needed and safely reuse cached results elsewhere. We give split uniform-convergence guarantees that size data needs separately for the critical and optional strata, clarify when passive i.i.d. sampling incurs a 1/p_c draw penalty, and show that if criticality can be detected pre-label or pre-compute, the label and compute budgets decouple from rarity. For contextual cache adjudication, we analyze an “exists-a-hit” reuse rule over k cached items with linear per-pair features and prove tight capacity scaling, with a reduction under block-separable features. For policy distillation into depth-D decision trees, we show a necessary logarithmic depth to reach a target value loss and provide imitation sample-complexity bounds that scale with 2^D up to log factors. Composed via a simple union bound, these pieces yield end-to-end reliability budgets and practical checklists for when to stratify, when reuse is statistically safe, and how deep an interpretable controller must be.