Can Generated Code Exercises Repair Small Model Blind Spots

This methodological proposal asks whether targeted generated code exercises can repair a small model's failures on naturally sourced tasks under a fixed data-production budget. It specifies an error-stratified comparison with targeted natural records, untargeted generated exercises, and a no-additional-data baseline, plus a sealed natural-task evaluation and cost accounting. A constructed task-mix example reverses a nominal route preference; its rates are arithmetic illustrations, not observed model outcomes. No model was trained for this article, no proposed comparison or sealed evaluation was run, no measured cost or repair effect is claimed, and the paper has not undergone peer review. ChatGPT and Codex assisted source review, drafting and editing, arithmetic checks, schematic figure preparation, and formatting.