The problem
On 30 October, X’s open-source account announced that the Under the Hood report now lists which of your posts received a visibility-limiting label, not just that some did (X Open Source). The ranking weights X publishes in its open-source repository invite a different reading: a conversion table, in which a copy-link share is worth about 30 likes, a reply on a post from someone you mutually follow about the same, a report wipes out hundreds of likes, and bookmarks are worth nothing.
The arithmetic behind that table is real. It comes straight from the defaults in param.rs in xai-org/x-algorithm. The reading is wrong, and the repository says so itself. The useful question for anyone who posts on X is what these two releases actually let you check, and what they still hide.
What I concluded
X has split its transparency into two halves. The ranking code publishes the formula: which actions the model predicts and how much each prediction counts. Under the Hood publishes the outcome for your account: which labels touched your posts last month. Neither half tells you why a given post got low reach, and treating the weights as an exchange rate between actions leads you to chase the wrong behaviour.
The 30 October change is a deliberate product choice, not a half-finished feature. Posts are listed only when a label is rare on your account. When a label covers a large share of your posts, X shows a percentage and no list, because at that point the label is about the account, not the post.
Constraints
- The weights multiply probabilities, not counts. The score is the sum of weight × predicted probability of each action (README). The README’s own example: “1 report cancels out 468 likes” is incorrect, because a report is far less likely than a like in the first place. The code comment on the report weight says its baseline probability is more than 1000x lower than a like (param.rs).
- Ranking and visibility are separate systems. The README describes the scoring pipeline and the visibility filtering as different stages. Under the Hood reports labels; it does not report the out-of-network discount, author-diversity decay or the 48-hour age filter, all of which can lower reach without any label.
- Some of the rules are withheld on purpose. The README says Grox prompts and some botmaker rules are not published, to prevent gaming. On the same day as the Under the Hood update, Nikita Bier described anti-spam work as “a bandaid for 3-12 months – until the adversaries find a workaround” (X).
- Not everyone gets a report. Under the Hood needs an account at least a year old and at least 10 posts in the prior month (X Open Source, September); the code carries the same thresholds (underTheHoodReport.User.strato).
The options, and what I traded away
| How to read the release | What you get | What it costs |
|---|---|---|
| Weights as a price list (“copy link = 30 likes”) | A simple rule to optimise | Wrong by construction; pushes you toward asking for shares and replies instead of earning them |
| Weights as a description of what the model is trained to predict | A sense of which reactions the ranking values once they happen | No per-post diagnosis; you still don’t know your own probabilities |
| Under the Hood as a complete reach report | One place to look | It only covers labels; ranking-driven low reach never shows up there |
| Weights plus Under the Hood, checked monthly | Separates “labelled” from “just ranked low” | Monthly lag, eligibility limits, and the withheld rules stay invisible |
I would take the last row. It gives up the comfort of a single number in exchange for a check that can actually come back with an answer.
Evidence
The weights. In the default production parameters: like 0.66, reply 5.0, repost 1.0, share 2.0, share via DM 5.0, share via copy link 20.0, quote 5.0, follow author 4.0, dwell 0.05. Negative: not interested −52.272, block −31.2, mute −58.8, report −234.0 (param.rs). Divide by the like weight and you get the ratios the price-list reading uses: copy link about 30x a like (20 / 0.66), a reply about 7.6x (5 / 0.66), a report about 355x (234 / 0.66). The README’s warning uses 468 for the report-to-like ratio; that does not match the current defaults (234 / 468 = 0.5, not today’s like weight of 0.66), and the README does not say which values the 468 came from. When an eligible original post comes from an author the viewer mutually follows, the weight on the viewer’s predicted probability of replying rises from 5.0 to 20.0 (a 15.0 boost), about 30x the like weight; that is the arithmetic behind “a mutual’s reply = 30 likes” (weights.rs, BIDIRECTIONAL_BOOST_CHANGE.md).
Bookmarks really are absent. The weighted-score function has no bookmark term (scoring.rs). That says nothing about whether bookmarks are useful to you; the 30 October announcement itself was bookmarked 10,526 times as of 2 November.
The weights move. The mutual-follow reply boost was A/B tested from 10 July at 5, 10, 15 and 20, rolled out at 20 on 13 July, and cut to 15 on 24 July after World Cup–related feedback (BIDIRECTIONAL_BOOST_CHANGE.md). The README logs further changes on 8 October (the ranking model predicts time spent better) and 13 October (retrieval is also trained to surface posts people read all the way through). Anyone who memorised July’s table is already out of date.
Only home-timeline engagement counts. The param.rs comments say engagement counts only on posts served in Home Timeline, and that mass reporting mainly affects recommendations for users similar to the reporters.
How Under the Hood lists posts. In the published code, with its default parameters, legal takedowns are always listed. For other labels, the specific posts are listed when 3 or fewer carry the label; above that, they are listed only while the label is on 5% of your posts or less (the list is hidden once the share is above 5%), and at most the 50 most recent posts per label are shown (underTheHoodReport.User.strato). X’s post says “fewer than 5%”; the code’s test is “5% or fewer”, so a label on exactly 5% of your posts is still listed, and the report’s own note says “5% or fewer”. The 3 / 5% / 50 values are defaults in the published feature-switch file (params.strato); production can set different values, so read them as the code’s defaults, not confirmed production settings. The change landed in the repository on 30 October UTC, the same day as the announcement.
Where this stops being true
- The Under the Hood thresholds (3 / 5% / 50) are feature-switch defaults in the published
params.strato, and production may use different values. The README says cron scripts set the defaults inparam.rsto the primary production values, but it makes no such statement about these report parameters. - The README says experiments running at a notable share of traffic (e.g. 10% or more) aim to be visible in the repo, which leaves room for smaller ones that are not.
- If X publishes the withheld botmaker rules, or starts reporting ranking adjustments per account, the “two halves” picture changes.
What I would do differently
If I ran a monthly check on an X account, it would be this. From the 10th of the month, open Under the Hood. That date comes from the published report code, not from an X statement: minDaysAfterMonthEnd = 10 means last month’s report only replaces the one before it from the 10th (UTC). If a label lists specific posts, read those posts first; that is the only place X tells you what triggered it. If a label shows only a percentage, treat it as an account-level problem rather than hunting for one bad post. If there are no labels and reach still dropped, stop looking for a penalty: that is ranking, and the honest levers are the ones the weights describe, which are posts people reply to, send to someone, and keep reading. Writing “please share” does not change the probability the model predicts. It is the probability that carries the weight.