A field guide, read straight from the source code
No folklore. This is what X’s open-sourced ranking code literally optimizes for — and a tool to score your draft against it.
The transformer predicts the probability you’ll act on a post; a weighted sum sets its rank. Raise the positive actions, avoid the negatives, survive the filters. That’s it. ranking_scorer.rs
01 The Analyzer
Paste a tweet. Every point added or subtracted is tied to a real mechanic in the code. It estimates the shape of a post the algorithm rewards — not the exact model weights (those aren’t in the repo).
02 What the code actually rewards
There are separate prediction heads for reply, quote, profile-click, and three kinds of share. A like is the cheapest action a person takes — the model rewards content that makes you want to weigh in or send it on.
● 19 discrete prediction heads ·recsys_model.py
There’s a dwell head, a continuous dwell-time head, and not-dwelled as a penalty. Making someone stop scrolling counts even if they never tap a thing — and scrolling straight past you actively hurts.
ranking_scorer.rs
Followers see you via Thunder (in-network). Virality means reaching strangers via Phoenix retrieval (out-of-network) — and OON content is multiplied by a penalty below 1, so it must outscore posts from accounts they already chose to follow.
● effective_oon_weight ·ranking_scorer.rs
block, mute, report, not-interested are predicted and subtracted. Rage-bait shown to strangers is self-defeating: it farms the exact buttons that lower your score. A spam/“banger” classifier labels low-quality bait before it’s even scored.
● negative_sum · Grox classifiersOld posts get filtered out of candidate generation, and the in-network store only keeps recent posts. Early engagement is what pulls you into more feeds — a slow start rarely gets a second life.
● AgeFilter · Thunder retentionAn author-diversity multiplier attenuates your 2nd, 3rd, Nth post in a single viewer’s feed — decaying by position. Posting twenty times doesn’t stack. One excellent post beats five mediocre ones.
● diversity_multiplier ·ranking_scorer.rs
Conversation dedup surfaces only your highest-scored post per thread, per viewer. Your 8-tweet thread isn’t 8 shots on goal — put your most quotable, most dwell-inducing line in the opening tweet.
● DedupConversationFilterImages and the transcribed audio of your video are embedded (multimodal V4/V5). Video that clears a watch threshold earns a dedicated “qualified view” weight. What you say on camera is indexed — give it a verbal hook.
● Grox embedder · vqv weightReplies run through their own ranking pipeline. A genuinely sharp reply under a large, relevant account ranks in its context, drives profile-clicks (a positive head) and follows — riding distribution you don’t have yet. Low-effort replies get flagged as spam.
● Grox reply-ranking pipeline03 The Playbook
Take a stance or ask a real question. Reply & quote are weighted far above likes.
Stop the scroll before anything else. Dwell is rewarded; not-dwelled is punished.
Coherent topics let retrieval route you to a real audience and build a recognizable embedding.
Strong-and-clean beats inflammatory-and-reportable. The negatives subtract.
Post at peak audience time. The age filter gives you hours, not days.
Per-author decay means your best single post is your real currency.
Conversation dedup surfaces one post — make the hook tweet the strongest.
Borrow bigger audiences while you’re small. Reply-ranking + profile-click is the hack.
It’s embedded and transcribed. Give video a verbal hook that clears the watch threshold.