X has open-sourced the core code for its "For You" recommendation system, publishing not just ranking logic but also weight parameters and the label structure tied to visibility limits. For people who publish on the platform, the move gives an official reference point for two questions that were usually answered through trial and error: what kinds of posts travel farther, and whether an account or post has been tagged in ways that restrict distribution.
On Thursday, X released the core code for the "For You" timeline on GitHub under the Apache v2 license in a repository called xai-org/x-algorithm. Keith Coleman, X's vice president of product, told TechCrunch: 「你会拿到拉取帖子、为用户排序并组装信息流的核心排序代码，其中一些系统，比如排序器和打分逻辑，你甚至可以在公司之外自己运行。」 Alongside the code release, the company also began piloting an "Under the Hood" page for users who want to download their own account data and check whether visibility-related labels have been applied.
How the feed is assembled
Based on the published code described in the report, X first gathers candidate posts from across the platform when a user opens the app. Half come from accounts the user follows. The other half come from machine-learning retrieval systems designed to surface posts from accounts the user may not know but may still care about.
A model called Phoenix then predicts how the user is likely to react to each post. Those predicted responses are grouped into five buckets:
- engagement, including likes, replies, reposts, and shares
- clicks, including clicks into the post, the author's profile, or linked content
- attention signals, such as dwell time and video watch progress
- following the author
- negative feedback, including "not interested," mute, block, and report
Each predicted action gets a probability score. That score is multiplied by its assigned weight, and the combined total becomes the post's overall score in the recommendation pipeline.
Even then, a high score does not guarantee that a post will appear without adjustment. The code includes three additional gates. Posts from the same author are progressively discounted starting from the second item, which discourages flooding the feed. Posts from unfamiliar accounts receive an overall discount, which leaves follower base size relevant. New authors, on the other hand, get a one-time boost designed to push them toward a target exposure level.
Why the share metric drew the most attention
The parameter file at home-mixer/params/param.rs quickly became the focal point of the release because it spells out how much different user actions are worth inside the scoring system.
Tech blogger Alex Finn, after reading 300,000 lines of code, posted a summary that reached 570,000 views. His main takeaway was that one share-link click carries roughly the same weight as 40 likes. That finding gave creators a concrete parameter-based reason to think less about optimizing for likes and more about creating posts people want to share.
The report says that interpretation fits the broader design of the algorithm. Likes sit near the low end of positive feedback value, while conversation and sharing carry more weight. Negative signals are much more costly. A single mute or block can require more than 100 likes to offset, and those penalties accumulate into an account reputation score that can affect distribution of later posts as well.
For creators, the practical reading is straightforward. Content that prompts reposting has an edge over content that only collects passive approval. Replies matter because conversation is valued more than one-sided applause. And external links in the main body of a post are not ideal if the goal is reach; the report notes that placing them in the comments may align better with a system that does not favor sending users off-platform.
Users can now check whether visibility labels were applied
The tool may be even more immediately useful than the code for ordinary users. X is testing an "Under the Hood" page in settings that allows eligible accounts to download a JSON file showing whether the account or its posts received "limited visibility" labels during the past month.
Eligibility in the pilot is limited. The report says the feature is being randomly offered first to users whose accounts are at least one year old and who made more than 10 posts over the past month. Broader access is expected to come later. The self-check entry point listed in the report is x.com/i/under_the_hood.
The repository also lays out the machinery behind the labeling system. It includes classifiers for spam, adult content, graphic violence, and hate symbols, as well as models aimed at detecting coordinated inauthentic behavior and an account reputation framework. Visibility decisions fall into three levels: normal display, placement behind a warning screen, or outright discard.
One ruleset applies only when content is being recommended to non-followers. That means a post may still be visible to an account's followers while disappearing from recommendation surfaces. The report presents that as a technical explanation for why some users have felt they were being "shadow limited" even though their followers could still see their posts.
Some users have already posted screenshots or data excerpts showing that two of their posts were marked NSFW and hidden from non-followers and minors. The report also notes that the label list does not include a category specifically targeting political content. For users who cannot read JSON easily, the company suggests handing the file to a large language model and asking it to interpret the data against the repository.
More complete than X's earlier algorithm releases
This is not X's first algorithm release. The company published code in 2023, but that version had gone three years without updates and no longer matched the live system. A rewrite released in January this year exposed the scoring framework but not the specific weights, which meant outside observers could see the weighted-sum structure without knowing the value assigned to each action.
This latest release fills in the missing pieces by adding weight parameters, filtering logic, and model configuration. According to the report, outside researchers can now use the open-source code to train and run the scoring system independently, something the previous releases did not enable.
Some parts remain closed. Prompting used by Grok for content-violation judgments, parts of the anti-spam rule set, and advertising-system code were not included. The stated reason is to keep malicious actors from using the code to evade moderation.
The self-audit tool is also still limited to a small randomized group, so most users cannot yet download their own data. Even with those limits, the report argues that X has gone farther than Meta, TikTok, and YouTube by putting both recommendation logic and visibility-restriction logic on the table at the same time, rather than only describing systems in research papers.
For creators, the long-running uncertainty around how to write for reach and how to tell whether distribution is being curtailed has become easier to inspect. One answer now sits in a public parameter file. The other is turning into a downloadable JSON record.