I trained a 15M "decision-only" model two ways — linear recurrence vs Transformer attention. Accuracy was a coin flip; latency and training cost weren't.[R]

Inspired by TypeSafe's Jev (a non-generative model that only outputs bounded decisions — choices, probabilities, calibrated confidence), I built two small decision models end-to-end on identical budgets and tasks to answer one question: which substrate makes the better Jev?