Reading the public code
What X's open-source algorithm can and cannot tell you about reach
The repository is detailed enough to kill several ranking myths, but it cannot explain the exact reach of one post. Knowing where the evidence stops is what keeps a useful code reading from turning into another fake algorithm hack.
The short answer
X's open-source algorithm can show how the For You feed gathers candidates, which actions the ranking model predicts, how those predictions are combined, which filters exist, and how documented changes move through experiments.
It cannot show the exact score your post received, the full live configuration for every viewer, every system used in production, or the single reason one post stopped at a particular number of impressions.
The right way to use the repository is to separate three things:
- what the published code directly shows;
- what X says about production defaults and experiments;
- what you infer and still need to test.
The source for this article is X's official For You feed repository, including its sections on experiments, unpublished files, deployment code, and the Under the Hood tool.
Questions the repository can answer
The August 2026 release provides unusually specific answers to broad questions about recommendation mechanics.
Where do For You posts come from?
Thunder provides recent posts from followed accounts. Phoenix retrieval and SimClusters find candidates from accounts the viewer does not follow.
What does the ranking model predict?
The public action list includes engagement, clicks, dwell and active time, following the author, and negative feedback such as mute, block, report, or not interested.
How are the predictions used?
RankingScorer multiplies each predicted value by a configuration weight, adds the results, and applies further adjustments before selection and reranking.
Which posts leave before scoring?
The published filters cover duplicates, posts older than 48 hours, already-seen content, viewer blocks and mutes, muted keywords, inaccessible subscriber posts, and several recommendation-specific cases.
Can a documented change be followed over time?
Yes. The bidirectional-follow document records an A/B test, a broad rollout, and a later parameter change. Our mutual follower boost analysis translates that history without turning the parameter values into impression promises.
Questions it cannot answer from the outside
The repository cannot reconstruct one post's complete production journey after the fact.
It does not give you:
- the candidate pools built for every viewer who might have seen your post;
- each viewer's predicted action values;
- the final score and reranked position for your candidate;
- the exact experiment assignment for every request;
- a causal explanation for the post's public metrics;
- a guaranteed format, topic, or posting time that will win future ranking decisions.
Public account analytics show outcomes. Public source code shows possible mechanisms. Neither alone proves which mechanism caused one outcome.
X documents gaps in the public code
The README says a limited set of files is not currently published because exposing them could make abuse easier. Its examples include specific Grox prompts and some BotMaker rules.
X also says the repository focuses on code that affects For You visibility. Some build, deployment, and internal infrastructure pieces are absent even where the decision logic is inspectable.
This does not make the repository useless. It makes "the entire production algorithm is open source" too broad a claim.
Defaults are not the same as every live request
X says many tunable values come from a configuration system. Cron jobs update defaults in the repository to match primary production values, while experiments can send different configurations to portions of timeline traffic.
The repository aims to expose experiments running at a notable share of traffic, using 10% or more as an example. That is an aim described by X, not proof that every experiment or assignment appears instantly.
When someone publishes a screenshot of one parameter and calls it "the algorithm," check the file date, surrounding code, experiment notes, and the distinction between a default and a viewer's actual request.
Four rules for reading algorithm claims
Check the source and date
Use the official repository and record the commit or update date. Old explanations of the previous Twitter codebase may describe a system that no longer matches the 2026 For You pipeline.
Read the unit, not only the number
A weight on predicted probability is not a reward per public engagement. The ranking weights guide explains this mistake in detail.
Keep ranking and visibility separate
A low rank and a visibility decision can both reduce exposure, but they use different inputs and services. The visibility filtering guide shows the two paths.
Mark inference as inference
If the code uses dwell predictions, you can infer that attention matters to the score. You cannot jump from that fact to "long posts always rank higher." The second sentence needs evidence the repository does not provide.
Combine public mechanics with private account evidence
The repository is good at generating better questions. Your account history is better at showing which ideas deserve another attempt.
PilotMyX does not decode X's hidden ranking system. It helps you inspect posts and replies, compare periods and content groups, preserve the examples behind the totals, and let an AI agent plan an approved publishing test from that evidence.
Use a simple evidence ladder:
| Level | What you know |
|---|---|
| Published fact | The behavior or parameter appears in X's official code or documentation |
| Reasonable inference | The mechanism suggests a possible content implication |
| Account observation | Your own comparable posts repeatedly show a pattern |
| Tested decision | You changed one thing, scheduled enough repetitions, and measured the result |
Most algorithm hacks stop at inference and pretend they reached the last row. Keeping the levels separate produces less exciting promises and far better decisions.
Use your own account