01How the For You feed is built
Every refresh builds a new feed: find candidates from several sources, filter, score each post with a model, then run a safety check. These 7 steps follow the code’s actual order.
- 1Read the viewerRecent engagement history, follows, blocks, mutes, muted keywords, already-seen posts
- 2Find candidatesAccounts you follow (Thunder) + accounts you don’t (Phoenix retrieval, SimClusters, popular posts)
- 3HydrateText, media, author details & labels, language, engagement counts, subscription
- 4Pre-filterOlder than 48 hours, duplicates, your own posts, blocked/muted, already seen…
- 5ScorePhoenix predicts the chance of 20+ actions per post, then weights are summed
- 6Pick top 50Sort by score, keep the top 50; about 35 posts are returned
- 7Visibility checkLabels decide: allow, notice, interstitial or drop
The pipeline in code
PhoenixCandidatePipeline finds, filters, scores and selects posts; ForYouCandidatePipeline then blends in ads and Who to Follow. Each request keeps the top 50 and returns about 35 posts.
02Where your posts get distributed from
Your post reaches people through two channels: followers (in-network) and non-followers (out-of-network). Both are ranked by the same model, with different rules.
In-network: your followers
Thunder keeps your followed accounts’ recent posts in memory and returns them on refresh — your followers are your most reliable audience.
Out-of-network: non-followers
Three sources: Phoenix retrieval (nearest posts in embedding space), SimClusters (clusters of who engages with what) and a shared popular-posts list. This is where new followers come from.
Replies and reposts don’t travel out-of-network
OONRetweetReplyFilter removes replies and reposts from accounts the viewer doesn’t follow. To reach strangers, post originals — replies and reposts won’t be recommended to them.
03The scoring formula
The model doesn’t look at a post’s like count. It predicts the probability that you will take each action, multiplies by a weight and adds them up.
Final Score = Σ ( weighti × P(actioni) )P = the model’s predicted probability that this viewer takes the actionFinal score = Σ (weight × predicted probability)
Phoenix predicts 20+ actions — like, reply, repost, quote, share, click, dwell, follow, not interested, block, report… Positive actions get positive weights, negative ones negative weights; the sum is then offset to stay positive.
Key point: weights multiply predicted probabilities, not real counts — so you can’t convert weights into “N likes cancel one report”.
04Action weights: what counts most
Production defaults from the code. “≈ likes” shows the weight relative to a like’s weight — a weight comparison, not a count conversion.
Positive actions
+2030× +2030× +57.6× +57.6× +57.6× +46.1× +23.0× +11.5× +0.661.0× +0.40.61× +0.30.45× +0.20.30× +0.050.08× +0.050.08× +0.020.03×Negative actions
-234355× -58.889× -52.27279× -31.247× -0.020.03×Columns: weight, and the weight relative to a like (≈ likes). Click any bar to open its line in param.rs.
05What makes reach take off
The weights and rules, translated into actions. Each point is backed by code.
Content that sparks replies and quotes weighs 7.6× a like
Replies and quotes weigh 5; a like weighs 0.66. Posts people want to answer or quote with their own take get recommended far more than posts that only earn a like.
Replies from mutual follows on your originals: weight 5 → 20
The July 2026 bidirectional follow boost adds 15 to the reply weight when the viewer and author follow each other, on original posts. Genuine mutual relationships with people in your niche are the most direct amplifier.
Make people want to follow you
Follow the author weighs 4 — 6× a like. A consistent niche and a clear profile raise the predicted chance that people follow you.
Get people to click and read to the end
Time spent after clicking weighs 0.4 per second. In October 2026 ranking got better at predicting time spent, and retrieval started favoring posts people read to the end. Clickbait that people leave immediately triggers the negative “not dwelled” signal.
Small accounts get a chance: the new-author boost
Posts can be lifted toward positions 15–16 when the author has ≤ 50,000 followers, the post is an original (not a reply or repost), it’s under 2 hours old and the author’s impressions are below the threshold (200). Thompson sampling picks which ones get tried.
Takeaway: post originals as a new account and stay around for the first 2 hours to reply.
The window is only 48 hours
AgeFilter removes posts older than 48 hours from For You. Post when your audience is active — the first hours matter most.
06What gets your reach limited
In code, limits happen at two layers: ranking pulled down by negative signals, and visibility rules that drop or hide posts based on labels.
Negative feedback carries huge weights
Report -234, mute -58.8, not interested -52.272, block -31.2. These actions are rare (a report’s base rate is 1000× lower than a like), so the weights are large: if the model predicts people like this will be annoyed, your rank with them drops fast.
Account labels: dropped from recommendations
Visibility rules drop posts from For You recommendations when the author has labels such as NSFW (high recall / precision), spam (high recall), compromised, read-only, impersonation, NSFW avatar/banner, abusive (high recall) or do_not_amplify.
This is what people call a shadowban: followers still see you in Following, but you stop being recommended to strangers.
Blocks and reports relative to likes
Agatha batch jobs compute blocks, reports and spam reports relative to the likes your posts receive (reports over 180 days) and label accounts from these ratios.
Credibility of who engages with you
user-cred-v2 runs PageRank over follow and engagement edges (engagement over 7 days) to give each account a credibility score. Engagement from low-credibility accounts (pods, sock puppets) doesn’t make you look trustworthy.
Inauthentic behavior patterns
The bdsm model reads an account’s sequence of actions over time to spot inauthentic or abusive behavior; abuse-enforcement-service then labels, challenges or suspends based on the scores. Mass following, mass replies and automation are high-risk.
If the parent is dropped, so are you
AncillaryVFFilter drops a post when its parent, quoted post or reposted post was dropped. Replying to or quoting risky content can take you down with it.
07Follows, unfollows, deleting, posting and replying: what the code says
Checked one by one against the open-source code: what helps, what isn’t counted and what hurts. Anything not in the code is marked “not in the public code” — no guessing. We’ll update as soon as more is published.
Mutual follows / mass following
Mutual follows get an explicit ranking boost; follows are also read one by one by the behavior model, where FollowBot is one of 8 detectors.
- Helps: when you and the author follow each other, the weight on “you’ll reply to their original post” gets +15 on top of 5 (BidirectionalFollowReplyWeightBoost = 15).
- Detection: bdsm reads an account’s recent action sequence with timestamp-based positions, which the README says capture “burstiness, mechanical cadence”; one head is FollowBot. SERVER_PROFILE_FOLLOW is in its action map.
- Enforcement: FollowBot’s thresholds are redacted (9.99 placeholders); production values are not published.
- Exemptions: accounts under 48 hours old, and accounts shown the login starter pack within 24 hours, skip FollowBot enforcement (both on by default).
Frequent unfollowing
Unfollow is in the behavior model’s action map, but the public code has no rule or threshold specific to unfollowing.
- SERVER_PROFILE_UNFOLLOW and CLIENT_TWEET_UNFOLLOW_AUTHOR are in the action map and are fed to the model with all other actions.
- How the model weighs them, and whether any unfollow count triggers anything, is not public; we’ll update as soon as it is.
Not in the public source yet — we’ll update this as soon as it’s published.
Deleting posts
The public code has no action or rule about deleting an original post.
- The model’s 196 actions include UNFAV / UNREPLY / UNQUOTE / UNRETWEET, but no “delete original post”.
- No ranking, filtering or enforcement rule references deletions. Whether deleting matters is not public; we’ll update as soon as it is.
Not in the public source yet — we’ll update this as soon as it’s published.
Posting a lot
Ranking doesn’t cap how much an author posts, but duplicate text, AI slop and gibberish get labeled, and posts from SpamHighRecall accounts are dropped from For You.
- Ranking: author diversity is off in value_model’s default config (enable_author_diversity: false), so there is no per-author decay.
- Duplicate text: scheduled jobs find clusters of identical posts and label them COPYPASTA_SPAM; one job matches CJK characters (bbq_duplicate_text_characters_cjk).
- Account labels: llm_slop_user → SpamHighRecall for 7 days; SpamEmbeddingMajorityPoster → 30 days; post-level gibberish_post → 30 days.
- Effect: when the author has SpamHighRecall, their posts are dropped from For You for followers and non-followers alike.
- TweetSpamBot: the code calls it a disabled head; it can only bounce to a captcha when ≥ 90% of posts come from one official client (off by default). Legit services posting via third-party API apps never pass this gate.
Replying a lot
Replies are among the highest-weighted positive actions, but fast reply spam, AI slop replies and duplicate replies each have a label.
- Helps: ReplyWeight = 5 (a like is 0.66), +15 more for mutual follows.
- fast_reply_spam_post → the post gets SpamHighRecall for 30 days; llm_slop_post → RiskyHighVizReply for 30 days.
- Account level: bb1_reply_spam_bot_bounce → SpamHighRecall for 30 days + liveness check; bb1_reply_spam_bot_suspend → suspension.
- Duplicate replies: word- and CJK-character jobs for replies also apply COPYPASTA_SPAM.
- Reach: OONRetweetReplyFilter drops replies and reposts from accounts the viewer doesn’t follow (by default), so your replies mainly reach your followers.
Engagement pods / link-sharing groups
Official comment: navigating straight to a post (e.g. coordinated in a group chat) has no ranking impact; only engagement on posts served in Home Timeline counts.
- From the comment: engagement counts only “on a post served in Home Timeline”.
- The behavior model also has LikeBot, RTBot, EngagementAmplifier and MultiActionBot heads, with thresholds redacted.
Pinned or posted low-quality links
If your pinned post links to a BAD / LOW_QUALITY URL, every follow you make triggers a check that labels you SPAM_HIGH_RECALL for a week.
- Rule 20732 (event: follow): pinned post with a BAD or LOW_QUALITY URL → SPAM_HIGH_RECALL for one week; the support note says “Restricting notifications”.
- Rule 6754 (event: url_event): a redirect chain containing a LOW_QUALITY URL whose final URL isn’t GOOD/WHITELIST → the post gets SPAM_HIGH_RECALL.
Large or high-credibility accounts
Enforcement rules first check follower count and credibility score and skip accounts above the threshold; the published value (12.34) is a placeholder.
- Both user- and post-level rules start with high_follower_count and pagerank_skipped skips; comments say production uses different values.
- The duplicate-text and low-quality-link rules also skip IsHighPageRankUser and IsUserGrayVerified accounts.
- user-cred-v2 runs PageRank over follow and engagement edges to produce each account’s credibility score.
08Why circle boosts work: each step mapped to the code
Circle boosts aren’t about inflating numbers — they build the relationships the algorithm actually uses. Every step links to the source.
- 1
Follow → you enter their in-network source
In the codeThunder keeps recent posts in memory and returns those from accounts the viewer follows — the in-network source of For You.
In our circleCircle follow tasks are verified through the API, after which your new posts become in-network candidates for that member.
- 2
Followers see your originals without the out-of-network discount
In the codePosts from accounts the viewer doesn’t follow are multiplied by a factor below 1, as are replies and reposts from followed accounts.
In our circleThe more members follow you, the more people rank your originals at full score.
- 3
Follow back → +15 reply weight for mutuals
In the codeFor originals from mutual follows, the reply weight is 5 + 15 = 20. Changelog: launched July 2026, set to 15 on July 24.
In our circleOnce members follow each other, replies to each other’s originals seen in Home are weighted higher.
- 4
Repost → reaches the reposter’s followers
In the codeOONRetweetReplyFilter only drops reposts from accounts the viewer doesn’t follow, so a member’s repost is an in-network candidate for that member’s followers (scored with the discount above).
In our circleRepost tasks put your post in front of members’ followers — a path the code allows.
- 5
Follow edges feed account credibility
In the codeuser-cred-v2 runs PageRank over the follow graph and engagement edges to score each account; high-credibility accounts skip some enforcement rules.
In our circleCircle follows and engagements happen between real X accounts, so they enter this graph. Exact weights and thresholds are not public.
What circle boosts can’t do — and why we’re built this way
In the codeThe code is explicit: engaging with a post by opening it directly (e.g. coordinated in a group chat) has no ranking impact on that post.
Our designSo the value of circle boosts isn’t inflating one post — it’s the relationships the code really uses: follows, mutual follows and repost distribution. Once connected, your future posts show up in members’ Home feeds, where their engagement does count. That’s also why we dropped like tasks, cap daily tasks and points, and offer no automation.
09Common misconceptions
“One report cancels 468 likes”? Wrong
The code comments explicitly correct this: weights multiply your predicted probability, not counts. Reports weigh a lot because they are so rare.
“Engagement pods or mass reporting control the feed”? Mostly no
Only engagement on posts served in Home Timeline counts. Navigating to a post directly (e.g. coordinated in a group chat) has no ranking impact, and since recommendations are personalized, bad-actor reports mostly affect people similar to them.
#L389-401“More likes is always better”? Not really
A like weighs only 0.66. Replies, quotes, shares, follows and time spent matter far more.
10Practical checklist
- Post originals — replies and reposts don’t reach non-followersWhy ↓
- Leave something people want to reply toWhy ↓
- Build real mutual follows in your nicheWhy ↓
- Write things worth saving and sharingWhy ↓
- Stay around for the first 2 hoursWhy ↓
- Post when readers are active — 48 hours is the windowWhy ↓
- No spam, flame wars or hard sellingWhy ↓
- No pods, sock puppets or automationWhy ↓
- Don’t reply to or quote risky contentWhy ↓
- Check labels with Under the Hood and our limit checkWhy ↓
11Algorithm update timeline
Changes published in the repository README and docs.
- Bidirectional follow boost A/B test starts (5/10/15/20)BIDIRECTIONAL_BOOST_CHANGE.md
- Boost value 20 rolled out broadlyBIDIRECTIONAL_BOOST_CHANGE.md
- Boost value lowered to 15param.rs
#L404-409 - Weights, visibility filtering and Phoenix training code publishedREADME
#L57-72 - Weights clarified (not count conversions)README
#L50-55 - Under the Hood shows legal restrictionsREADME
#L46-48 - Ranking predicts time spent betterREADME
#L42-44 - Retrieval trained for read-throughREADME
#L38-40 - Brazil election filter removedREADME
#L34-36
12Source index
- home-mixer/params/param.rs — 权重与参数
- xai-value-model/scoring.rs — 打分公式
- xai-value-model/weights.rs — 互关加成
- home-mixer/scorers/author_cold_start.rs — 新作者加速
- home-mixer/candidate_pipeline/phoenix_candidate_pipeline.rs — 流水线
- home-mixer/filters/ — 过滤器
- visibility-filtering/rules/ — 可见性规则
- agatha/ — 拉黑 / 举报比例标签
- user-cred-v2/ — 账号可信度
- docs/BIDIRECTIONAL_BOOST_CHANGE.md — 互关加成变更记录
This page explains the open-source code; actual production values can be changed by experiments and configuration. Not affiliated with X.