X publishes the For You weights, and a way to check whether your account is flagged

X said on Thursday that it is publishing the source code for the For You timeline on GitHub under the Apache v2 license, expanded to include model configuration, filters and core ranking system details, along with the parameters used to weight different signals ^[1][2][3]. For anyone whose reach depends on that feed, the change that matters is not the gesture but the artefacts: weights you can read, and, according to X's VP of Product Keith Coleman, components such as the ranker and the score that can be run outside the company ^[4].

The release is roughly 10 to 15 times larger than X's previous open source codebase ^[5]. Coleman says external researchers who previewed the repository got the score, the numerical value calculated for every post, running outside X, which he describes as a major milestone for the company's transparency efforts ^[6]. That is the load-bearing claim for practitioners. A score you can reproduce offline is a bench you can put candidate posts through, which moves argument about distribution from anecdote toward something closer to measurement.

The account-level piece is thinner. A transparency tool is rolling out to an "Under the Hood" page in settings, letting users who have posted 10 or more times in the past month download aggregate stats as a JSON file showing whether any labels have been applied to their account or posts over the past calendar month ^[7][8]. It starts as a pilot limited to a test group of accounts at least a year old, before a broader rollout ^[9]. So the initial audience is the intersection of two filters: at least ten posts in the last month and at least twelve months of account age ^[1]. And because the window is the past calendar month, an older label will not show up in a given download ^[2]. X's suggested route for non-technical users is to drop the file into an LLM, point it at the GitHub repo, and ask for an interpretation ^[10].

There is a deliberate gap. Systems that use Grok to predict whether a post might violate a rule are not included, which the company says protects it from bad actors who could work around its rules to flood the network with spam ^[11]. If a post is suppressed on that path, the published code will not explain it.

X will accept pull requests, which its engineers will consider incorporating, though not all will make the cut ^[12]. Coleman's framing is that the algorithm could be not just visible to the public but written in part by it ^[13]. The stated purpose is to address continuing concerns about the algorithm's influence on politics, elections and misinformation ^[14], in a company whose predecessor faced Republican allegations in Congress of shadowbanning conservative posts, which Twitter consistently denied ^[15], and which is now owned by a trillionaire who helped elect President Trump ^[16]. Coleman says the goal is code that can be audited and critiqued so people trust the system is fair ^[17].

Watch three things: whether the pilot's age and volume gates come down, whether the label names in the JSON map cleanly onto filters visible in the repository, and whether any outside pull request actually ships.