starl3xx.fun / algo
X Algo Watch
Report
Every Sunday: what actually moved in the open-source ranking code behind X’s For You feed (xai-org/x-algorithm), diffed against last week’s mirror and translated for people who post. Read from the code, not from folklore. One post, Sunday night, Central; quiet weeks get one standing rule instead.
User actions that increase your reach
param.rs at 3b6919f, 7 November 2026Read a multiple as a ratio between two weights, not as a promise about reach: each one multiplies a predicted probability rather than a count, so a reply weighted 7.1 times a like is not a reply being worth 7.1 likes. These are defaults in open-source code, and the repo says itself that they may not match production. Since issue 005 a separate ranking service returns the final score, and once it has loaded its config it sums these same weights from its own copy of these nine defaults. That copy matches this one on seven of them; it still has the like at 0.5 and the share at 2.0. Without that config, or when the call to it fails, home-mixer’s own sum of these weights is served.
And two things you cannot earn
×0.625 is 0.25 + 0.75 × 0.5, from AuthorDiversityDecay at 0.5 and AuthorDiversityFloor at 0.25 in vm-ranker/params.rs. The ranking service applies them once it has loaded its config, which its open-source startup flag leaves off, and not at all when home-mixer’s call to it fails; whether X runs it that way is not in the repo. Until issue 005 this page said ×0.5 for every extra post, but only the part above the floor halves, and the formula is the same at every commit this series has read. 48h is MAX_POST_AGE in home-mixer/params/config.rs, written there as 48 * 60 * 60.
Updates
This week · 8 November 2026Issue № 011 (commit 3b6919f)A second rise for the like, the video open weighted at zero, a quarter more weight on time spent on a post, time away from Home as a new Phoenix input, and a Grok weight at zero0.5 → 0.66video open 0.07 → 0.02dwell time 0.004 → 0.006not interested -47.52 → -52.272How it works
- Diff the recorded commit against HEAD.
- Parse every default in
param.rs,config.rsand the ranking service’sparams.rsat both ends, and compare by name, so no number is retyped. - Write what moved, and what it means for posting.
- Check every figure back against the diff.
- starl3xx reviews. Nothing ships unread.
