Phoenix: Candidate-Isolated Transformer Ranking in X's Recommendation System

Technical analysis of X's Phoenix recommendation system, open-sourced January 19, 2026. This paper examines the novel candidate isolation attention pattern where post candidates cannot attend to each other during transformer inference. Key contributions include comparison of 2023 Heavy Ranker vs 2026 Phoenix architectures, analysis of candidate isolation design and its implications for caching, consistency, and fairness, enumeration of 19 engagement signals predicted by Phoenix, comparison to other production systems (YouTube, TikTok/Monolith, Instagram), and discussion of transparency limitations.