X to publish 10-15x more of its feed ranking code, including the visibility filters

X plans to put the ranking stack behind its "For You" feed on GitHub under the Apache v2 license, in a drop the company says is roughly 10 to 15 times larger than its previous public release ^[1][3]. For anyone whose distribution depends on that feed, the useful part is not the gesture but the artifact: some of the code is meant to be executable outside the company, which replaces inference from impression counts with something you can run ^[4].

The repository is described as containing model configurations, filters, and the signal scoring parameters that set post order in the personalized feed ^[2]. Keith Coleman, quoted by TechCrunch, said the release covers the core ranking code that pulls posts, ranks them for a specific user and assembles the feed, and that the systems filtering potentially rule-violating content will be visible; he added that some of those systems, specifically the ranker and the scorer, can be run outside X ^[4].

Alongside the code, X is preparing an in-app transparency feature called "Under the Hood" in settings, intended to show whether ranking systems affected the visibility of an account or of individual posts ^[5]. Accounts with at least ten posts in the past month will be able to download a JSON summary of the labels applied to the profile or its posts during the previous calendar month ^[6]. The pilot starts with a limited set of accounts at least a year old ^[7]. Because each export covers one calendar month, reconstructing a year of label history would take twelve separate pulls ^[16]. X says no special technical skill is needed, and suggests handing the JSON to a language model together with the GitHub repository address to get an explanation of how the algorithms may have affected reach ^[8]. That is a workflow recommendation, not a guarantee of accuracy.

Two limits are worth reading closely. Outside developers can file pull requests, but X engineers decide what gets merged into the algorithm ^[9]. And X will not release the tooling that uses Grok to predict likely rule violations in posts, arguing that publishing it would help bad actors evade enforcement and push spam ^[10]. So the filters you can read are the ones X considers safe to expose, and a locally run scorer will not include the withheld classification path ^[4][10].

The stated goal, per Coleman, is that any member of the public can assess how posts spread, judge whether the playing field is level, and criticize the system so it can be fixed ^[11]. In practice the immediate value is narrower: users get a way to check whether impressions were throttled or posts hidden from search without notice, the pattern usually called shadowbanning ^[12]. That question sits inside a long-running argument about how X ranking affects political content, elections and misinformation ^[13]. It also arrives against a thinner disclosure baseline; since going private, X has published far less about audience, revenue, growth and government takedown requests ^[14].

Things to watch: whether the pilot widens past year-old, high-frequency accounts ^[7][6]; whether any external pull request is actually merged ^[9]; whether the label names in the JSON map cleanly onto filters in the repository, which is the test of whether the two releases are one system or two ^[2][6]. On the money side, X is also replacing Revenue Sharing with Original Content Rewards, with new applicants eligible from 8 September under a Premium subscription and set thresholds ^[15].