A focused fine-tune of Gemma 4 12B on verifiable Python coding data — every training example's reasoning leads to code that actually passed its tests. The result reasons in the open (edge cases, complexity, approach) and then emits a clean, runnable solution.