x algorithm series
Twitter out of network: how your posts reach people who don't follow you
11 min read
Updated 21 October 2026. We wrote this in May 2026 from the repo as it was then. It claimed twelve candidate sources, a mixture-of-experts lane running in parallel, a Grox topic layer feeding every feed, and an OON_WEIGHT_FACTOR in a file called oon_scorer.rs with a hidden value. Today’s code wires nine sources into the post pipeline, the mixture-of-experts and semantic-ID sources are off by default, there is no oon_scorer.rs, and the out-of-network discount has a visible code default of 0.75. This update rebuilds the article from the current code and adds the new-author boost, the out-of-network filters and the extra visibility rules for recommendations.
One change since this refresh matters for reach outside your network: from 22 October, when a feed has more than 2,800 candidates, every source is trimmed by the same share instead of cutting the last source in line. Our October 2026 algorithm update explains it with a before and after chart.
A creator with 5,000 followers can only reach those 5,000 people through the in-network lane. Everything beyond that runs through what the code calls out-of-network (OON) candidates: posts from accounts the viewer does not follow. So how does Twitter out of network discovery actually work, and what decides whether a stranger sees your post?
The short answer from the root README.md: in-network posts come from Thunder, out-of-network posts come from Phoenix retrieval, SimClusters and a shared list of popular posts, and both are ranked together by the same model. The longer answer is a set of sources, filters and multipliers, each with a number you can read in the code. This article walks them in order. It pairs with the series hub and with how Phoenix represents authors and posts.
The candidate sources in today’s code
The folder home-mixer/sources/ holds 20 source files, but most belong to other surfaces: ads, Who to Follow, prompts and push-to-home are added by the blending pipeline, and others serve the Following feed or internal scoring endpoints. The For You post pipeline, home-mixer/candidate_pipeline/phoenix_candidate_pipeline.rs, wires nine. Their defaults sit in home-mixer/params/param.rs, a file the repo says is mirrored from production defaults.
The nine sources in the For You post pipeline
Source | What it finds | On by default? |
|---|---|---|
| Recent posts from accounts you follow (in-network), up to 1,200 | Yes |
| Out-of-network posts nearest your history, up to 1,000, plus up to 200 “cold” results | Yes ( |
| Out-of-network posts similar to ones you engaged with, up to 800 | Yes ( |
| A shared list of popular recent posts, up to 500 | Yes ( |
| Phoenix retrieval limited to a topic | Only when you open a topic feed |
| A mixture-of-experts retrieval variant | No ( |
| Posts sharing semantic-ID prefixes with ones you engaged with | No ( |
| An older mixing service | No ( |
| Reuses already-ranked posts within a session | Only when cached posts exist |
Source: home-mixer/candidate_pipeline/phoenix_candidate_pipeline.rs and home-mixer/params/param.rs, read 21 October 2026

The README’s candidate source, retrieval index and ranking tables, github.com/xai-org/x-algorithm, captured 21 October 2026.
So, for a creator in October 2026, three lanes reach strangers on the main For You feed: Phoenix retrieval, SimClusters and popular posts. The README notes that experiments can switch values for a share of traffic, so “off by default” means off for most viewers, not for every viewer.
Phoenix retrieval: two towers, one index
Phoenix retrieval is a two-tower model. The full mechanics are in our Phoenix embedding article; the short version from phoenix/README.md:
User tower: encodes your recent engagement history (home-mixer sends up to 1,024 actions,
MaxSeqLengthRetrieval) plus coarse profile features. There is no learned ID for you; you are represented by what you interacted with.Candidate tower: represents each post by its semantic ID (a content code derived from its text and media) plus its hashed author ID.
Search: the top posts by dot product between your vector and a precomputed index come back as candidates.
Retrieval is trained with favorites as the positive signal. In plain terms, “near your history” means “close to posts that people with histories like yours went on to like”.
The retrieval index: how a post gets in, and what keeps it out
The README’s retrieval index table names two components. phoenix-rankall/ maintains the index Phoenix retrieval queries, and phoenix-rankall-strato/ decides which index a post belongs in, “consulting visibility filtering first.”
The routing logic in phoenix-rankall-strato/columns/phoenix_rank_all/phoenixRankAllCandidateProcessor.strato shows what that means:
A new original post goes into a post-creation index when it is published.
When it receives a like, it is added to a “1fav” index, a topic index and, at 32 likes or more, a “32fav” index.
Reposts and community posts are skipped. Replies only enter separate reply indexes, and only after a like.
Any post that visibility filtering would drop at the home recommendations level, or that is flagged as adult content, is skipped (adult video has its own index).
Our read: a post that visibility filtering would hide from non-followers never becomes a Phoenix retrieval candidate for strangers at all. The filter does not just trim it later; it keeps it out of the index.
The cold lane for new posts
phoenix_source.rs asks for up to 200 extra “cold” results on top of the main 1,000 (PhoenixColdStartMaxResults), served as ForYouPhoenixRetrievalCold. Phoenix’s own code defines a home “cold” slice in phoenix/xrex/data/retrieval_dataset.py and phoenix/xrex/data/cold_pool_filter.py: posts with fewer than 8 likes and fewer than 500 views, and by the training config’s default no older than 2 hours. The serving code adds that slice to home retrieval.
The client that links the 200-slot request to that slice is not in the repo, so treat the connection as our read. If it holds, fresh posts from smaller accounts get a reserved place in retrieval during their first couple of hours.
SimClusters and popular posts
SimClusters was added to the repo on 13 August 2026 as an extra out-of-network source in simclusters/. The README describes it as clustering accounts and posts by who engages with what. In home-mixer/sources/simclusters_source.rs:
Seeds are posts you engaged with, explicitly or implicitly, from the last 90 days.
For each seed it finds similar posts by cluster embedding and keeps those scoring above 0.5.
Candidates must be 48 hours old or less (
SimclustersMaxCandidateAgeHours), with up to 800 returned.OONNsfwSimclustersFilterdrops SimClusters posts from adult-flagged authors you do not follow.
Popular posts is a shared list, not a personal one. home-mixer/util/popular_posts.rs and home-mixer/popular_posts_job.rs build it from original posts by the most-followed active accounts, from the last 24 hours, ranked by views adjusted for age (an 8-hour half-life), with at most 5 posts per author and 500 in total. Our read: for most small and mid-sized creators, this lane is not theirs to win.
Filters that only hit out-of-network posts
Before scoring, every candidate goes through the pre-scoring filters, including the 48-hour AgeFilter. Two of them target out-of-network posts specifically.
OONRetweetReplyFilterremoves reposts and replies from accounts the viewer does not follow, plus any reply whose parent is missing. A switch to allow some out-of-network replies from Phoenix retrieval exists (EnablePhoenixOonReplies) but is false by default.NewUserMinEngagementFilteris off by default. When on, it applies only to viewers whose account is under 30 minutes old (or who returned within the last 30 minutes) and drops out-of-network posts below a like threshold.
Practical consequence: on the For You feed, your replies and reposts do not travel to people who don’t follow you. Your original posts do.
Extra visibility rules for recommendations
After ranking, visibility filtering checks each post. In-network posts are checked at the TimelineHome level; out-of-network posts at TimelineHomeRecommendations, which starts with every home rule and adds more, all of which can only drop a post. The lists are in visibility-filtering/rules/levels.rs:
Account labels: several adult-content labels, spam high recall, compromised, read-only, impersonation, abusive high recall and do-not-amplify.
Post labels: DMCA or geo-restricted media, adult media, do-not-amplify, malicious URL, spam high recall, abuse insults and the Brazil election legal drop.
The README puts it plainly: the same post is allowed to a follower. One example shows how a label gets there. In abuse-enforcement-service/service-lib/rules/enforcement_user.yaml, an account whose model score carries llm_slop_user receives the SpamHighRecall label for 30 days, and that label drops its posts from recommendations to non-followers. You can check which labels apply to your own account with the Under the Hood tool, and run a quick outside check with our shadowban checker.
The out-of-network discount
Once Phoenix has predicted the 25 actions for every candidate, the score is a weighted sum: each weight in param.rs multiplies a predicted probability, then offset_score keeps every score positive. The repo’s own warning applies: weights multiply predicted probabilities, not counts, so reading the report weight as “1 report cancels out 468 likes” is wrong, and the comments in xai-value-model/scoring.rs note that a report’s baseline probability is more than 1000x lower than a like.
Then the vm-ranker service applies three adjustments. The out-of-network one is effective_oon_weight:
0.75 for posts from accounts the viewer does not follow (
OonWeightFactorinvm-ranker/params.rs).The same 0.75 for replies and reposts from accounts the viewer does follow (
EnableOonRescoreForInNetworkRepliesRetweets, true by default; the rule isoon_appliesinxai-value-model/scoring.rs).0.5 when the viewer is browsing a topic feed (
TopicOonWeightFactor).A factor of 0.00001 for brand-new viewers also exists, but its age threshold is 0 seconds, so it never triggers by default.
Two caveats. These are code defaults in a file that, unlike param.rs, is not marked as synced from production, so the live value is not confirmed. And the README only says the factor is “below 1”. The arithmetic, if 0.75 holds: an out-of-network post needs a base score about a third higher than an equal in-network original to land level with it.
The other two adjustments matter for out-of-network reach too. Author diversity multiplies each extra post from the same author in one feed load by a decaying factor (code defaults: ×1, ×0.625, ×0.4375, down to a floor of ×0.25). Then a determinantal point process reranker in vm-ranker/ trades a little score for less similarity between neighbouring posts. Whether that reranker runs in production depends on startup flags that are not in the repo. Our read: a post that looks like ten others in the list is the easiest one to drop.
The new-author boost
The one adjustment that points the other way is in home-mixer/scorers/author_cold_start.rs, with its settings in param.rs. On each feed request, at most one post can be lifted. To be eligible, a post must:
be an original post, not a reply or repost;
come from an author with 50,000 followers or fewer (
ColdStartFollowerCap);be 2 hours old or less (
ColdStartMaxPostAgeSecs7,200);have fewer than 200 home timeline impressions (
ColdStartImpressionThreshold);already sit above the bottom 3% of scored posts (
LowImpressionsMaxPositionRatio0.97).
Among eligible posts, the code uses Thompson sampling: it draws from a Beta(0.75 + likes, 49.25 + impressions minus likes) distribution for each, keeps the top 2 draws, and picks the one with the higher model score. That post’s score is raised to at least the 16th-best score in the list (ColdStartSlotMin 15, ColdStartSlotMax 16). It does not apply on topic feeds.

New-author boost settings in home-mixer/params/param.rs, github.com/xai-org/x-algorithm, captured 21 October 2026.
The prior works out to a 1.5% like rate. Our read: a small account’s fresh post that earns likes from its first viewers at a better rate than that has a real chance of being the one post lifted in a given feed load. The boost works for in-network and out-of-network candidates alike; it just matters most for accounts that have few followers to begin with.
An illustration, not code
Picture an applied machine learning creator with 4,000 followers who posts an original, specific observation from running a model in production. Here is how the code paths above could play out:
In its first two hours it is eligible for the cold retrieval slice and the new-author boost.
A first like puts it in the main retrieval index, where viewers whose histories are full of similar technical posts can retrieve it.
SimClusters can surface it to people who engaged with posts that sit in the same clusters.
For each stranger it reaches, its score is multiplied by 0.75 and then has to hold up against everything else in their list.
A vague, broad question on the same subject goes through the same gates. Our read is that it fares worse: its content code sits near a great many similar posts, so it is close to everyone’s history and especially close to nobody’s. None of this is measured in the repo; it is how we read the mechanics.
What to do with this
Facts first, then our read.
Originals carry discovery. Out-of-network replies and reposts are filtered out of For You, and in-network replies and reposts take the 0.75 discount too.
The first two hours are special. The new-author boost and, as we read it, the cold retrieval slice both work on posts up to 2 hours old. Every candidate is gone after 48 hours (see 48 hours in Thunder).
Make the subject obvious. Retrieval represents a post by its semantic ID and its author. Our read: posts with a clear, recognisable subject sit near the histories of people who care about that subject.
Stay clear of labels. The recommendation-only visibility rules can remove a post from strangers’ feeds entirely, and keep it out of the retrieval index. Check Under the Hood if reach to non-followers suddenly drops.
Do not chase the discount. 0.75 is the same for everyone. What differs between creators is the score before it applies.
Ads and modules take slots too. See the ways Twitter fills your feed slots.
VoiceMoat’s view, and it is our read rather than something the code states, is that consistent, recognisable content from a real person tends to earn the positive actions Phoenix predicts (replies, dwell, shares, follows) and avoid the negative ones (not interested, mute, block, report). That matters more out of network, where every post starts at ×0.75. VoiceMoat is a Personal Brand OS for LinkedIn and Twitter: audenAI, the brain inside it, learns from your own posts, and VoiceMoat checks drafts for Style and Stance, schedules and publishes them, and shows the analytics afterwards. It does not change how Twitter ranks anything. For a comparison with other tools, see VoiceMoat vs Tweet Hunter, Typefully and Hypefury, and for how reach builds over time, what actually compounds.
How does Twitter show my posts to people who don't follow me?
Through three out-of-network sources that are on by default: Phoenix retrieval (a two-tower model that finds posts near the viewer's engagement history), SimClusters (posts similar to ones the viewer engaged with) and a shared list of popular posts. Those candidates are ranked by the same Phoenix model as posts from accounts the viewer follows.
Is there a penalty for out-of-network posts on Twitter?
Yes, a multiplier. In vm-ranker/params.rs the out-of-network factor defaults to 0.75, and the same factor applies to replies and reposts from accounts the viewer follows. It is a code default in a file not marked as synced to production, so the live value is not confirmed; the README only says the factor is below 1.
Do replies and reposts reach non-followers on the For You feed?
No, by default. OONRetweetReplyFilter removes reposts and replies from accounts the viewer does not follow, and the switch that would allow some out-of-network replies is off. Original posts are what travel to strangers.
Do small accounts get a boost on Twitter?
Yes. On each feed request, at most one post can be lifted to about the 16th slot if it is an original post, the author has 50,000 followers or fewer, the post is 2 hours old or less and it has fewer than 200 home timeline impressions. The pick uses Thompson sampling on its like rate.
Why might my posts stop reaching non-followers?
Visibility filtering has extra drop-only rules for recommendations, covering account and post labels such as spam high recall, do-not-amplify, malicious URL and adult content. A post with one of those labels can still reach followers but is dropped for non-followers and kept out of the retrieval index. The Under the Hood tool shows labels on your account.
This article was drafted with AI assistance and human-edited by Prateek Singh. All technical claims are sourced to the xai-org/x-algorithm repository; file paths are cited inline.
More in the series

Twitter Algorithm
The Twitter algorithm in 2026: how the open-source ranker works and why your voice matters
Twitter's For You feed runs on open-source code in the xai-org/x-algorithm repo on GitHub, first published in January 2026 and updated almost every weekday since 13 August. A transformer called Phoenix predicts 25 actions for each post, published weights turn them into one score, and author diversity, an out-of-network discount, a new-author boost and a DPP reranker follow. This hub walks the current code in plain English and explains why consistent, recognisable content still matters for creators.
14 min read

Twitter Algorithm
How the Twitter Phoenix algorithm sees creators: hashed authors, semantic IDs and viewer history
Phoenix, the model behind the Twitter For You feed, never reads your "voice". It reads hashed author IDs, semantic IDs that describe each post, and each viewer's last thousand or so actions. This refresh walks the production code published since August 2026 and labels clearly where our creator advice is interpretation rather than code fact.
12 min read
The whole series
The Twitter/X algorithm in May 2026: all ten articles