X has published a fuller look at how its For You feed evaluates posts, after a major Aug. 13 update to the open-source x-algorithm repository added key recommendation parameters, ranking weights and visibility systems.
The update offers direct clues to questions creators have debated for years: which interactions matter most, whether dwell time really boosts distribution, how content from unfollowed accounts enters the feed, and whether account-level credibility signals exist.
How For You builds a recommendation set
According to the architecture X made public, the system rebuilds recommendations each time a user opens For You. Candidate posts come from two main sources.
One is In-Network, where Thunder pulls in recent posts from accounts a user already follows. The other is Out-of-Network, where Phoenix Retrieval and SimClusters look for posts from accounts the user does not follow but may still find relevant. Both streams are then handed to Phoenix for ranking.
X's published documentation says the recommendation stack works in two stages: Retrieval first narrows down a pool of millions of possible items, then Ranking uses a Transformer model to order the remaining candidates. Phoenix does not simply ask which post is hottest overall. It looks at a user's recent interaction history and predicts what that user is likely to do after seeing each candidate post.
Those predictions span actions such as Like, Repost, Reply and Click. The model then combines them with different weights to produce a final score.
Replies outrank likes in the published coefficients
One of the clearest disclosures in the repository is a set of ranking weights. In the current configuration, Favorite, meaning a like, carries a weight of 0.5. Repost is set at 1.0. Reply is set at 5.0.
On that basis, a reply carries 10 times the weight of a like and five times the weight of a repost in the final combination step. Quote is also set at 5.0. A standard Share is 2.0, while Share via DM is 5.0.
The highest figure in the published list is Share via Copy Link at 20.0. That makes it 40 times the like coefficient of 0.5 and four times the reply coefficient of 5.0.
X's code does not mean one copied link literally equals 40 likes. The system first estimates the probability that a user will take each action, then multiplies those probabilities by action-specific weights. X also says the weights reflect both ranking value and how often those behaviors typically happen on the platform.
In practical terms, if Phoenix believes a user is highly likely to copy and share a link after seeing a post, that prediction carries substantial value in the ranking formula.
Mutual follows add an extra boost to replies
X also exposed a parameter tied directly to social relationships: BidirectionalFollowReplyWeightBoost = 15.0.
That means if two accounts follow each other, a reply between them receives an extra 15.0 boost. The signal shows X does not treat every reply as identical. The relationship between the two accounts can affect how the interaction is scored.
Negative feedback carries steep penalties
The repository also cuts against the long-running idea that any engagement is good engagement. Phoenix models negative reactions as well as positive ones.
In the currently published parameters, Not Interested has a weight of -43.2. Block Author is -31.2. Mute Author is -58.8. Report falls as low as -234.0.
So the ranking system is not only asking whether a post can trigger replies or shares. It is also asking whether the post is likely to push a user toward muting, blocking, dismissing or reporting the author. A controversial post may generate replies, quotes and shares, but it can still lose ground if the model expects strong negative responses at the same time.
Dwell time is part of the model, but not in a simple way
Dwell time has long been treated as a key distribution metric by many creators. X's published examples show Phoenix does model attention behavior, and dwell appears as one of the action types used to learn user preferences from historical interactions.
Still, the released ranking parameters do not support a simple rule that longer reading time always leads to more reach. The general DwellWeight is listed at 0.0. At the same time, ContDwellTimeWeight has a positive weight, while NotDwelledWeight is negative.
That suggests X incorporates attention and dwell behavior in more than one form. It is not a straightforward leaderboard of seconds spent reading a post.
Posts from unfollowed accounts are actively sourced, with a discount factor
X's architecture also explains why users regularly see posts from accounts they have never followed. Phoenix Retrieval searches a large content pool for relevant candidates, while SimClusters looks at relationships between users and content to surface material the user may care about.
X labels this category Out-of-Network. It does not need a follow relationship to recommend a post.
Those posts are not on equal footing with content from followed accounts, though. The code includes an Out-of-Network discount factor, meaning posts from unfamiliar accounts may receive an extra scoring adjustment. X is looking for new content outside a user's network, but it appears to impose a higher bar before that content is shown.
Cold-start support exists for smaller and newer accounts
The recommendation system includes a Cold Start or New-Author Boost mechanism for posts with low exposure and for smaller creators. The published parameters include ColdStartImpressionThreshold = 1000, ColdStartFollowerCap = 1000, ColdStartMaxPostAgeSecs = 86400, ColdStartSlotMin = 15 and ColdStartSlotMax = 16.
Under those settings, a post may be pulled toward around the 15th to 16th slot if it has fewer than 1,000 impressions, comes from an author with no more than 1,000 followers, and was published within 24 hours.
That is one of the more concrete signals for small creators in the public code. It suggests X does not reserve recommendation space only for accounts that already have momentum. The system keeps room to test lower-exposure content and decide whether it deserves broader distribution.
Author diversity limits repeat appearances in one request
X also addresses a common concern about whether posting too often causes a creator to compete against their own content. At least within a single For You ranking request, the platform does not want one author to dominate the feed.
The repository includes an Author Diversity mechanism. After the first candidate from a given author, later candidates are multiplied by a declining factor to prevent one account from filling the recommendation page with consecutive posts.
X disclosed this as a per-request diversity mechanism, not as a daily posting penalty formula. The repository does not answer how many posts per day is optimal.
The public repository does not support the "golden 30 minutes" rule
Social media advice often claims the first five minutes, first 30 minutes or first hour determines whether a post will break out. The currently public repository does not provide enough evidence for those specific time rules.
What X does confirm is that its pre-scoring filters remove candidate posts older than 48 hours. The system also filters out duplicate content, the user's own posts, posts from blocked or muted accounts, content matching muted keywords, and posts the user has already seen.
At a minimum, that means posts do not stay in the standard recommendation candidate pool forever.
No module called Shadowban, but visibility filtering can limit reach
X's public materials do not describe a system explicitly called Shadowban. But the repository does show a Visibility Filtering system that can decide whether content is shown normally, dropped outright, or placed behind an INTERSTITIAL that requires extra user action to view.
The filtering layer considers a user's own block and mute settings, along with labels generated for accounts and posts by other systems. X also disclosed related systems including Agatha, BDSM, user-cred-v2, Botmaker and Scarecrow, which help generate signals and labels that can affect content visibility.
From a user perspective, that means a post may still exist while reaching a smaller audience through recommendation limits, even if X does not name the mechanism Shadowban.
Account-level scoring exists, but not as a simple hidden score
The code also points to an account-level scoring system. user-cred-v2 uses information such as the social graph to generate account-level scores. Agatha and BDSM also analyze accounts and user behavior to produce labels that may later affect visibility filtering.
That means X is not evaluating only the post itself. During Candidate Hydration, the system adds author information, account labels, language, engagement counts and subscription status.
Even so, the public repository does not support the simplified claim that every account has a single hidden credibility number from 0 to 100. The disclosed system is more complex than that.
Premium and external-link penalties remain unproven in the code
The same caution applies to two popular claims on X: that Premium subscribers receive direct traffic boosts, and that external links are automatically downranked.
Candidate Hydration does fetch subscription status. But knowing whether an account is subscribed is not the same thing as showing a fixed ranking boost for Premium users. Based on the currently public repository alone, there is not enough evidence to confirm claims that Premium gets a constant 2x or 4x traffic multiplier.
The code also does not provide enough support for the idea that any post containing an external URL is automatically hit with a penalty factor. Phoenix may predict link-related user behavior, but that is different from a blanket downranking rule for posts with outside links.
On the basis of the released code, neither claim can be treated as confirmed.