Before X can rank a post for someone, it needs to find that post as a candidate. The public “For You” code describes several paths into that pool. Following an author is one path; it is not the only one.
The distinction matters because discovery and ranking answer different questions. Retrieval asks which posts are worth considering. Ranking estimates how a particular reader might respond to those candidates. A post that enters the pool has an opportunity to compete, not a promise of a place in the feed.
Two broad paths into the pool
The repository’s pipeline overview describes Thunder as an in-network source: it keeps recent posts in memory and returns posts from accounts the viewer follows. That supplies material from an existing social graph.
For posts beyond that graph, the same overview describes Phoenix retrieval and SimClusters. Phoenix represents a viewer and posts as vectors and retrieves nearby candidates. SimClusters groups accounts and posts by engagement patterns, then uses those clusters to find candidates. A shared popular-post source is also described.
Think of the flow as: followed accounts → Thunder; other discovery → Phoenix, SimClusters and popular posts → candidate pool → ranking and later selection. This is a simplified map of the reviewed code, not a complete diagram of every X surface.
Similarity creates an opportunity, not a guarantee
Suppose, hypothetically, a reader often engages with explainers about blockchain security. A new post on that subject from an unfamiliar author might be retrieved as a candidate if it is represented as relevant to the reader. This is an illustration, not evidence that a specific post was retrieved, or that any particular topic signal alone caused it.
The Phoenix documentation distinguishes a retrieval model that narrows a large collection from a ranking model that scores a smaller set. The two-tower retrieval design makes representations for the viewer and for posts, then searches for candidates using similarity. That is a way to select possibilities, not a human-readable rule saying “show this topic.”
Why the source is not the outcome
Once retrieved, a candidate still passes through other stages. The overview describes filtering, scoring and additional selection or visibility checks. Being followed does not ensure a post survives these steps; being outside the reader’s network does not automatically exclude it.
For an author, the useful distinction is between eligibility for consideration and the eventual response of a real audience. A total impression count does not identify which retrieval source supplied a post or which later step determined its position. The code also does not give an author a universal recipe for entering a reader’s candidate pool.
Based on the public code snapshot pinned for this article. Actual production configurations, experiments and trained model weights are not established by this review. The discussion concerns the examined “For You” route, not every way to find a post on X.
