SARA (Scaling Articulated Rationales) is an industrial framework that leverages articulated user rationales (AURs) to improve recommendation systems. It builds a data engine to elicit and curate AURs, aligns a general-purpose MLLM into SARA-7B, and integrates generated rationales into production ranking. SARA outperforms strong MLLM baselines in offline evaluation and online A/B tests.