X has published a larger version of its For You ranking code and paired it with a new label transparency tool. If you build an audience on X, you now have more to inspect, but not the whole machine.
On August 13, X updated its public For You feed algorithm repository on GitHub, and this release goes deeper than the company's earlier transparency efforts. The code sits under the Apache License 2.0 and now includes configuration parameters, weights used to blend predicted user actions into a post score, visibility-filtering code, and systems that attach labels to accounts and posts.
That is the useful part. X's own repository says the For You feed pulls posts from accounts you follow and from accounts you don't, then ranks them with a transformer model. It also says the system uses signals from your recent engagement history, your follows, blocks, mutes, muted keywords, served posts and other inputs before deciding what should appear in the feed.
According to TechCrunch, X first open sourced part of its recommendation system in 2023, when the company was still trying to turn Elon Musk's transparency promise into something developers could actually inspect. That first release was easy to criticize because code without weights, live defaults, training data, or enforcement context can look more like a diagram than a working window into the platform. The August update is harder to dismiss. It names the machinery that shapes reach: Phoenix retrieval, Thunder, SimClusters, visibility filtering, safety labels and ranking scores.
You should still be careful with the word open.
Shadowban checking moves from rumor to labels #
The more practical change for ordinary users is X's new Under the Hood label transparency tool. The GitHub README says X is piloting a tool that lets people see aggregate statistics about visibility-impacting labels on their accounts and posts. The serving code for that report is also in the repository, under an under-the-hood folder.
That doesn't prove every shadowban complaint true. It doesn't make every ranking decision visible in real time. But it does move the conversation away from pure guesswork. If your account has been tagged by systems that can limit visibility, X says the tool is meant to show those outcomes and let you match the labels back to the code.
Shadowbanning accusations have followed Twitter and then X for years. In 2023, TechCrunch reported that Musk had promised to address the lack of transparency around the practice, while former trust and safety executives warned that enforcement systems were layered, messy and difficult to explain cleanly to users. That history matters because users don't just want a policy page. They want to know whether their posts are being quietly narrowed after they hit publish.
The code is still not the whole platform #
X is also clear about what it isn't publishing. The repository says Grox prompts and some botmaker rules are being held back because public release could help people game the system. That is a fair concern. It is also the reason skeptics will keep arguing that the release is incomplete.
Both things can be true. You can give developers more than they had before and still keep the most abuse-sensitive rules out of view. Any large social platform has to make that trade. The difference here is that X is putting enough of the ranking and filtering stack in public for outside developers to test claims against code instead of treating every reach drop as folklore.
The timing also fits a wider run of algorithm changes at X. TechCrunch reported in July that head of product Nikita Bier said X had tweaked its algorithm to make posts from mutuals more visible after replies started feeling too much like a battleground among strangers. That is the piece founders and marketers should notice. The system is not fixed just because the code is public. It changes, and those changes can alter who sees you.
If you run a brand account, a founder account, or a media operation on X, this release gives you something better than growth-hacker superstition. You can inspect how ranking separates scoring from visibility filtering. You can see that negative signals such as blocks, mutes, reports and not interested actions sit inside the model's action set. You can also see why a post can score well and still fail to appear if visibility filtering drops it. Frankly, that is more useful than another thread promising the perfect posting time. The real lesson is simpler: X is turning part of its distribution system into public infrastructure, while keeping the parts most useful to spammers behind the curtain. Trust will depend on whether the Under the Hood tool expands beyond a pilot, and whether future code updates keep matching what actually runs in production.
Also read: Nvidia Bets Its Balance Sheet That $500 Billion in AI Chips Won't Age • Meta's Smart Glasses Trigger a Privacy Backlash It Can't Fix With Software • A Single ISP Nearly Broke Solana and Exposed Its Hidden Concentration Risk