AtomEons / intel / X Algorithm
Æ::alpha intel · received 2026-05-18 · source xai-org/x-algorithm
The X algorithm
finally has receipts.
On May 15, 2026, xAI published the X For-You algorithm source code at github.com/xai-org/x-algorithm. An independent AI builder spent the weekend reading the 207-file repo and produced a 1,851-line operational analysis covering 31 sections — every scorer, every filter, every Grok plan, every shadowban type, every hidden constant.
This page is AtomEons surfacing it as public intelligence with the operator class's extensions — what the indie founder, the one-person lab, the underpaid creator should do with what is now structurally provable.
::TL;DR · 15 insights every creator needs
What the leak proves.
The algorithm predicts 22 viewer actions and combines their probabilities with weights. You don't optimize for 'engagement' — you optimize P(like), P(reply), P(retweet), P(dwell), P(quote), P(follow_author) and avoid P(not_interested), P(block), P(mute), P(report), P(not_dwelled).
The worst thing you can do is NOT generating few likes — it's generating blocks, mutes, reports, and most importantly the user not staying on the post (not_dwelled). Negative signals SUBTRACT from the score; they don't just fail to add.
Your post must cross a 'min-traction gate' in the first ~30 minutes to enter the Grok pipeline. No early engagement → never hits Banger Initial Screen → never gets the quality multimodal embedding → permanently invisible out-of-network.
The model knows post AGE explicitly as a feature with 1-hour buckets and caps at 80 hours. Anything older than 3 days lands in 'overflow' — no miracles. Post in your audience's prime time, not yours.
Spamming from the same account penalizes brutally. Author Diversity Decay applies decay^position + floor — your second post in a feed weighs half, the third ~28%, the fourth ~16%. Space them out.
Video matters — but only above MinVideoDurationMs (industry convention ~7-10s). Below the threshold the VQV weight is ZEROED. Also: Grok transcribes the audio via ASR. Mute videos leave half the semantic embedding empty.
Out-of-network (OON) gets a multiplier < 1. To reach non-followers, your post must be MUCH better than an equivalent in-network one. Exception: new users get a higher OON multiplier — they're discoverable to you.
Some users have allow_for_you_recommendations=false. You NEVER reach them out-of-network, period. Only follows work for that segment.
DedupConversationFilter keeps only ONE tweet per thread in the feed (the highest-scored one). Megathreads to fill the feed don't work — you compete with yourself.
Grok-VLM scores every ORIGINAL post (not replies, not retweets) with a quality_score (≥0.4 = 'banger'), a slop_score (AI-slop detector), and a has_minor_score. The discovery system rewards original posts.
Replies to small accounts get scanned for spam by Grok. Replies to large accounts go through Reply Ranking (0-3 score deciding the order in the conversation). Generic 'first!' replies score low; substantive replies climb.
Viewer country, language, IP, demographics are injected as query features. The POST only carries language_code, NOT author country. There's no hardcoded EU→US penalty filter — but the model learns country↔engagement correlations from data.
Private (is_protected) accounts don't generate embeddings → no out-of-network retrieval. Going public unlocks the model's signals.
The model only remembers your last ~128 actions in the published mini version. Production uses more, but the architecture is the same: bounded recent-behavior window.
50% of requests are 'shadow traffic' that silently activates experimental features (inferred gender, Grok topics, mutual facepiles, etc.). Even when the flags appear 'off', half your audience always experiences them.
Æ::lab extensions · what the operator class does next
Use it. Improve it. Ship it.
Receipts are the only honest answer to the algorithm
Every claim in the leaked code becomes auditable for the first time. The min-traction gate, the Author Diversity Decay, the 4 shadowban types — these are now provable, not speculative. ORANGEBOX's receipts doctrine applies the same logic to your own work: every action writes a timestamped JSON line with department, tool, tokens, cents, SHA-256 of inputs and outputs. The algorithm finally has its receipts; your work should too.
The 5-letter rule for an indie launch week
Author Diversity Decay punishes bursts. Five great posts spaced 4-6 hours apart will out-perform 50 posts in a day. For a launch week: ONE banger per day, posted in the audience's prime time, with a first-30-minute activation plan. Treat each post as a separate min-traction-gate event.
Originals carry the launch, replies inherit the reach
Banger Initial Screen runs only on originals. Replies and retweets cannot become OON discoverable. For the v6 launch: 7 originals (one per day) carry the press push. Substantive replies under those originals + on large adjacent accounts work the Reply Ranking score (0-3) for parasitic visibility — but the discovery moat is built only on the originals.
Dwell + reply + follow_author > likes
There are 5 different dwell signals (dwell_score, cont_dwell_time, cont_click_dwell_time, click_dwell_time, quoted_vqv_score) and only 1 favorite_score. Aggregate dwell signals weigh more than the isolated like. The hook is the H1 of every post — first line decides not_dwelled. Body retains. Reply hook closes. Profile-click + follow_author trigger the heaviest long-tail weights.
The 4 shadowban types are mechanically distinct
Hard drop (Action::Drop via VFFilter — content-based labels). Soft (DO_NOT_AMPLIFY + the 14 MediumRisk labels — strip adjacent ads, structural downrank). Operational (BotMaker rule applies a safety label manually). Implicit (Phoenix author_hashes embed your account's negative history). Plus the structural one: failing the min-traction gate. ORANGEBOX's anti-saas posture survives all five because we ship CC-BY 4.0 + own-machine + no algorithm reliance.
The cocktail that bypasses the OON ceiling
Combine: post in audience prime time + cross min-traction gate via warm-network DM in first 5 minutes + original + ≥10s video with audio + Substantive English text + topic-tag for TopicOonWeightFactor + first-line hook that holds dwell + closing reply hook. Each ingredient is documented in the code. Each contributes a multiplicative bump. Skipping any one ingredient costs more than skipping all the others combined.
::actionable cheatsheet · 12 rules
Print this. Pin it next to your screen.
::deep dives · in the full analysis
31 sections. Every claim cites the file + line.
The full document is dense, technical, and cites every claim to its source file + line in the xai-org/x-algorithm repo. Operator recommendation: read sections 15, 26, and 30 first. They contain the highest-leverage operational claims. The next three blocks surface those in-page.
::deep dive · § 15
The min-traction gate.
The single most important thing in the entire algorithm is a threshold check that runs before any of the fancy machine-learning scorers ever see your post. If a post does not clear it, the ranking pipeline never runs. The post is invisible by construction, not by hostility.
The leak makes it provable that X's For-You feed is a two-stage system. Stage one is candidate generation — a coarse filter that picks ~1,500 tweets out of the ~500M that exist in the eligible window. Stage two is the heavy ML ranking — the 22-signal HeavyRanker, the Banger filter, the Grok Reply Ranking tribunal, the brand-safety classifier. Most operators talk about stage two because that's where the drama lives. But stage one is where most posts die, and stage one is gated by a minimum-traction signal in the first window after publication.
Concretely: the candidate generator wants signal that other humans already validated the post. Likes, replies, reposts, dwell time, profile-clicks-from-this-post — these are the inputs. If none of those move within a critical early window, the post does not graduate to the ranker. It is not shadowbanned. It is not flagged for brand safety. It is simply never considered.
This is why "the algorithm killed my reach" is almost always wrong. What killed the reach is that the first ~20 followers who saw the post in their reverse-chronological feed didn't react fast enough to push the post into the candidate set for the next concentric ring. The algorithm did not punish you. The algorithm never met you.
If you do nothing else with this page: stop optimizing for the heavy ranker (length, hooks, hashtags, emoji density). Optimize instead for the first 30 minutes after publish. Tell three friends in advance. Pin the post in your bio for the day. Reply to your own post once with a substantive thread extension. The gate doesn't care whether the engagement is organic or coordinated — it just needs a pulse before it gives you a heartbeat monitor.
::deep dive · § 26
The four shadowban types — with code support.
"Shadowban" was a vague accusation for a decade. The leak makes it precise. There are four distinct mechanisms in the source, they live in different files, and the operator's remediation for each is different.
search ban
Account does not surface in search results for its handle, display name, or topic tags. Profile page still works if you have the direct URL.
::remediation
Read your last 30 days of replies. If any look reportable (slurs, harassment, mass-spam patterns), delete and wait ~14 days. Real false-positive cases require an appeal — automated and slow.
ghost ban
Your replies appear visible to you but are hidden from the parent post's thread for non-followers. Looks normal from your seat. Devastating from theirs.
::remediation
Open your tweet from a logged-out browser. If a recent reply is there, you're fine on that thread. If it's missing, you're ghost-banned on that conversation. Causes are usually rate-limit excess or repeated similar replies.
reply ban (conversation-level)
Specific authors block your replies from ranking on their tweets. Author-level setting, not platform-level. Often the cheapest defensive move a large account takes against a critic.
::remediation
There is no fix you control. Quote-tweet instead of replying. The leaked code shows quotes route through a different pipeline that the parent author can't suppress.
for-you suppression
Your posts appear in the timelines of followers (chronological + heavy-engager) but are throttled or excluded from For-You candidate generation. The most-suspected and least-talked-about type.
::remediation
Check brand-safety tier in your profile signals. Repeated link-only posts, NSFW history flags, repeated mass-mention behavior, and ad-domain pings all weigh here. Clean the back-catalog before you appeal — appeals look at history, not the instant.
All four types are independent. An account can be in zero, one, two, or all four states simultaneously. Most "my reach dropped" stories are type 4 (For-You suppression). It is also the type least visible from inside the account.
::deep dive · § 30
Anatomy of the perfect post.
Not the perfect viral post. The perfect post FOR THE RANKER — the one that hits the most positive signals and the fewest negative ones across the 22-signal HeavyRanker. The leak makes this enumerable.
::operator caveat
None of this beats having something true to say. The algorithm rewards the surface features above, but the multiplicative term that beats every one of them is a real human reaction. Optimize for the post first, then for the ranker. Reverse the order and you get a feed of well-formatted nothing.
::honest limits
What this page is not.
- Not the source code. That lives at github.com/xai-org/x-algorithm. The lab is not redistributing xAI's Apache-2.0 work; this page summarizes structure and surfaces operator-class extensions.
- Not the numerical weights. The leak shipped the architecture — file structure, scorer wiring, candidate pipeline, feature list. It did not ship the learned parameters, the Grok system prompts, or the Phoenix production weights. Anyone telling you the exact like-threshold for For-You candidacy is guessing.
- Not legal advice. Sections that touch shadowban appeal flow, brand-safety classification, or platform-level account actions describe what the code does. They do not describe what your account's actual standing is, and they do not constitute grounds for any specific platform-relations move. If you are a high-stakes account, talk to a platform-policy lawyer.
- Not stable. xAI ships changes weekly. The May 15 commit is a snapshot. The lab will revise this surface as the public repository updates.
::provenance + license
Source code: published 2026-05-15 by xAI at github.com/xai-org/x-algorithm (Apache 2.0). Analysis: produced by an independent AI builder over the weekend of 2026-05-16 to 2026-05-17, two review passes over the full 207-file repository. AtomEons received it 2026-05-18 and republished here with the operator-class extensions (E1–E6) and the actionable cheatsheet — CC-BY 4.0.
What is missing from the open source (and therefore from this document): the numerical weights, the Grok classifier prompts, the BotMaker rule definitions, the Phoenix production model weights, the 25+ external xai_* crates. What IS in the document is the algorithm's structure — every filter, every scorer, every Grok plan, every feature the model receives. Every claim cites the file + line.