The xai-org/x-algorithm repository, which houses X's For You feed algorithm, had accumulated 32,609 GitHub stars by September 4, 2026, making it a trending project S¹. The Rust-based project asserts that it contains the fundamental logic determining the posts displayed in a user's timeline [S1, P2]. The current surge in popularity stems from two August updates that exposed the mechanics behind the feed's post scoring and filtering. These revelations alter our understanding of the system curating countless timelines, while also highlighting questions the code alone cannot resolve.

My read: This is the most detailed look at a major platform's recommendation algorithm I have seen in a public repository. The August 13 update, which swapped a demonstration model for actual production training code, is the real headline. But I am cautious about one thing: no independent party has verified that this code matches what X runs in production. The Brazil election filter is politically sensitive and worth watching, but it is also a compliance mechanism tied to electoral law, not necessarily censorship. The labeling systems, with internal names like scarecrow and agatha, are the part I would dig into first, because those labels determine whether a post lives or dies in the feed.

How the feed scores a post

Recommendation generation occurs across two phases [S1, P2]. Initially, the system collects potential updates from two distinct sources: posts from followed accounts (in-network) and content surfaced via machine learning (out-of-network) S¹. As of the August 13 revision, SimClusters was introduced as an additional out-of-network source, complementing the pre-existing Thunder and Phoenix systems S¹.

During the second phase, a transformer-based model evaluates and orders these candidates S¹. Maintainers incorporated configuration settings on August 13 that combine various predicted action metrics into one comprehensive ranking score S¹. Subsequent comments added on August 14 highlighted an important distinction: these weights are applied to the estimated likelihood of specific user behaviors rather than simple tally counts S¹. Essentially, the system estimates the probability of a user interacting with a post (like liking or sharing) and multiplies that chance by a specific weight. The aggregate of these weighted probabilities forms the final score. Consequently, an update with 50 interactions from highly probable users could outrank one with 500 interactions from unlikely users.

That same revision also substituted the Phoenix demo model with the genuine training scripts used for live production models, alongside code for generating synthetic data S¹. This change is significant because the repository now features the actual training pipeline rather than a simplified placeholder.

The labels that decide if a post survives

Alongside ranking, the August 13 revision introduced visibility filtering logic that dictates whether a post is displayed normally, hidden completely, or placed behind an interstitial warning screen requiring a click to view S¹.

Supporting this filtering mechanism is a collection of labeling frameworks S¹:

  • botmaker and botmaker-rules: apply rule-based labels to user profiles
  • scarecrow: assigns labels to content
  • agatha and bdsm: evaluate profiles across multiple metrics (these are internal codenames rather than literal descriptions)
  • user-cred-v2: calculates a user's credibility rating
  • media-model-proxy and clip: analyze visual and video content
  • abuse-enforcement-service: manages enforcement measures

These classifications directly influence the visibility filter. For instance, content from a profile flagged by botmaker could be suppressed, while media flagged by clip might trigger an interstitial. Additionally, the August 13 update brought an Under the Hood transparency tool that provides aggregate data on these labeling systems S¹, offering insight into the frequency of their application.

The Brazil election filter in the code

Within the codebase lies a Brazil2026ElectionFilter designed to suppress content from accounts flagged by Brazil's Electoral Court during the 2026 election cycle, except for users who actively follow the flagged accounts S¹. The list of affected accounts saw its latest update on August 27, 2026 S¹.

Functioning as a compliance tool aligned with Brazilian electoral regulations rather than a universal content rule, this filter exclusively targets accounts reported to the Electoral Court. While it allows followers of these accounts to continue viewing their content, non-followers will find that such posts vanish from their For You feeds throughout the election period.

What to do about it

If you build or study recommendation systems, the training code and synthetic data generation pipeline added on August 13 are the most useful artifacts here. You can clone the repo and inspect how a production feed model blends predicted action probabilities into a ranking score, which is a cleaner description of the mechanism than most academic papers provide.

For a social media manager at a news outlet in São Paulo, the Brazil2026ElectionFilter has a direct consequence. If your outlet's account gets reported to the Electoral Court during the 2026 election cycle, your posts will stop appearing in the For You feeds of users who do not already follow you. Your existing followers are unaffected, but your reach to new audiences through the algorithmic feed drops to zero. The practical check this week: search for your account name in the filter's account list in the repository, which was last updated August 27.

What we don't know yet

All assertions in this analysis derive from the repository's README and embedded code comments [S1, P2, P4]. There has been no external verification that this open-sourced code accurately reflects X's live production environment. While maintainers state it includes the core logic, there may be undocumented components.

The Brazil2026ElectionFilter prompts several unresolved questions, such as the procedure for adding accounts to the Electoral Court list, the update frequency, and the existence of an appeal mechanism. Although the list was updated on August 27, the schedule for subsequent revisions remains unknown.

While systems like agatha, bdsm, and user-cred-v2 are identified in the codebase, the README lacks comprehensive documentation on their scoring methodologies. Deciphering the practical implications of these dimensions necessitates a direct review of the model code.

Looking ahead, monitoring the repository's commit history will be crucial for tracking updates to the Brazil filter account list as the October 2026 election nears. We will observe whether the list expands and if the filter's parameters shift. If you want that follow-up in your inbox, subscribe and we will send it.

Sources: S1 — xai-org/x-algorithm: Algorithm powering the For You feed on X · P2 — README.md · P3 — brainsparker/youagent · P4 — README.md · P5 — Paper page - DarwinX: Evolving Agent Harnesses Through Natural Selecti

More from Not A Tech Guy

Generated from an audited evidence pack with primary-source research. Social-media items are discussion signals, not verified facts. Nothing here is financial, legal or medical advice.