A deep-dive into the public Phoenix ranking model â a Grok-based transformer that combines in-network and out-of-network candidate retrieval, multi-action engagement prediction, Grox content understanding, author diversity attenuation, and feed-ranking to decide what users see in their For You feed.
The X Tool's scoring model is based on publicly available X recommendation-system signals, including multi-action engagement prediction, negative feedback risk, content understanding, audience fit, candidate hydration, author diversity, and feed-ranking concepts described in the May 15, 2026 public xai-org/x-algorithm repository release.
Instead of only asking whether a post might get likes, The X Tool estimates whether a post is likely to earn valuable engagement, avoid negative feedback, match audience intent, and remain easy for recommendation systems to understand.
Thunder + Phoenix Retrieval
Real-time signal enrichment
Transformer-based engagement prediction
64-dim interest embedding
Quality gates + diversity rules