Score any tweet, or your draft, with X's own algorithm.
Bangermeter is a free Chrome extension. It puts a 0 to 100 score on every post in your X feed, and on your draft while you write it, using the real production weights X published on August 13, 2026, re-checked against X's source as of September 18, 2026. Every factor is cited to the line of source code. No folklore, no data collection, no server.
Add to Chrome · Free Read the code
Weights shown are X's, as of September 18, 2026.
What you get
A score badge on every tweet
As you scroll x.com, each post gets a badge with up to two 0 to 100 scores (50 = a typical median post). C is Content Appeal: what the algorithm would predict from the content alone. E is Earned Engagement: how X's weights value what the post actually earned per view. A ~ marks low-view posts where rates are statistically smoothed.
Click for the plain-English breakdown
Big score chips, human explanations, detected content signals with ▲/▼ direction, rescoring factors, freshness decay, and a collapsible "Show the math" with the full weight × probability table.
A live draft meter while you write
The compose box gets a meter that scores your draft in real time with ▲/▼ hints: asks a question ▲, three or more hashtags ▼, "like if" bait ▼.
All 26 heads in the popup
Every head with X's published value, its param.rs key, and whether Bangermeter measures it or estimates it. The same table as further down this page, one click away while you browse.
Special detections
Community-Noted posts (score suppressed ×0.5, per three causal studies), visibility-limited posts flagged qualitatively, and quote-tweets handled correctly: embedded media is not counted as the quoting post's own.
A mutual-follow toggle
Scores the post as if the viewer mutually follows the author, which lifts the reply coefficient from 5 to 20 (as of September 18, 2026). It is the most interesting switch in the popup, and it shows how much of the algorithm is about who is reading. Plus badge, draft-meter and out-of-network toggles. Settings sync through Chrome's own storage. That is the only permission it wants.
How the score is built
Every weight below is X's, read from home-mixer/params/param.rs, as of September 18, 2026 (upstream sync stamp 2026-09-18T16:21:20Z). A weight is a coefficient on a predicted probability, so a big number usually means a rare action. The eight that matter most:
| Action | Weight | What it means | In Bangermeter |
|---|---|---|---|
| Share via copy link | 20 | The largest positive coefficient X publishes. Also one of the rarest actions. | Estimated |
| Reply | 5 | Becomes 20 when you mutually follow the author of an original post. | Measured |
| Quote | 5 | Matches the reply coefficient. | Estimated |
| Follow author | 4 | A new follower earned from the post. | Estimated |
| Like | 0.5 | Small coefficient, by far the most common action. More expected score on a median post than anything above it. | Measured |
| Profile click | 0.0 | Zeroed by X. The 2023 table paid 12 for this. | Zeroed by X |
| Mute author | −58.8 | Nearly twice the coefficient of a block (−31.2). | Estimated |
| Report | −234 | The heaviest coefficient in the system. It pushes a post beneath every positive-scoring post in the feed. | Estimated |
All 26 weights X published
| Action | Weight | What it means | In Bangermeter |
|---|---|---|---|
| Report | −234 | The heaviest coefficient in the system, in either direction. | Estimated |
| Mute author | −58.8 | Muting carries nearly twice the coefficient of blocking. Most people assume the reverse. | Estimated |
| Not interested | −43.2 | The "show less often" menu item, in X's own words. | Estimated |
| Block author | −31.2 | Milder than a mute, and milder than most people expect. | Estimated |
| Share via copy link | 20 | The largest positive coefficient X publishes. Also one of the rarest actions, so read the trap below before converting it. | Estimated |
| Reply | 5 | Rises to 20 on an original post from someone you mutually follow. | Measured |
| Quote | 5 | The same coefficient as a reply. | Estimated |
| Share via DM | 5 | Forwarding to a friend. | Estimated |
| Follow author | 4 | Gaining a follower off the post. | Estimated |
| Share | 2 | Opening the share sheet and sharing. | Estimated |
| Repost | 1 | Twice the coefficient of a like, and about as common. | Measured |
| Like | 0.5 | Small coefficient, by far the most common action. It carries more expected score on a median post than anything above it. | Measured |
| Post click | 0.4 | Opening the post itself. | Estimated |
| Open link | 0.2 | Links ARE rewarded. Small, and the "link penalty" story is wrong. | Estimated |
| Video open | 0.07 | Starting the video. X raised it from 0.05 on August 25, 2026. | Estimated |
| Photo expand | 0.05 | Tapping an image to full size. | Estimated |
| Dwell (binary) | 0.05 | X turned it on August 25, 2026 (it was 0.0). It fires only when the viewer engaged and stayed at least 10 seconds. | Estimated |
| Quoted-post click | 0.05 | Clicking through to the post being quoted. | Estimated |
| Post unexplored | 0.02 | A novelty term, in-network posts only. | Viewer-specific |
| Not dwelled | −0.02 | The largest term on an ordinary post. Tiny coefficient, and it fires on most impressions. | Estimated |
| Dwell time | 0.004 | Paid per second of dwell. One of the largest terms an ordinary post earns. | Estimated |
| Profile click | 0.0 | Zeroed by X. The 2023 table paid 12 for this. It is now worth nothing. | Zeroed by X |
| Video quality view | 0.0 | Zeroed by X on August 25, 2026. It was 0.05. Its gate measures the clip's own length (over 10 seconds). Watch time does not enter it. | Zeroed by X |
| Quoted video quality view | 0.0 | Zeroed by X. | Zeroed by X |
| Click dwell time | 0.0 | Zeroed by X. | Zeroed by X |
| Active seconds (5m residual) | 0.0 | Zeroed by X. | Zeroed by X |
All 26 scoring heads, sorted by absolute weight. Every value is X's, read from home-mixer/params/param.rs as of September 18, 2026, and each one is asserted individually by the engine's 255 self-tests. Measured means Bangermeter reads the real count off the post. Estimated means the weight is X's and the probability is ours.
What happens after the weighted sum (all figures as of September 18, 2026). A mutual-follow reply adds +15, on original posts only, never on replies or reposts. Out-of-network multiplies by ×0.75, applied exactly once, including to in-network replies and reposts. Author diversity applies (1 − 0.25) × 0.5k + 0.25 for the k-th post from the same author. The second post from an author keeps 62.5% of its score. The third keeps 43.75%.
How the scale is built. The scorer adds an offset of 0.001 (as of September 18, 2026) to the weighted sum, then runs offset_score(), which compresses any net-negative combined score into a band strictly below every net-positive post. Sum the published constants as of September 18, 2026 and the positives come to 43.34 against 367.22 of negatives, a total range of 410.56. Those three totals are my arithmetic on X's constants. The offset and the compression step live in the scorer (ranking_scorer.rs), not in param.rs.
That is the real cost of a report. A reported post falls beneath every positive-scoring post in the feed, whatever else it earned.
What changed on August 13
Until that morning the only ranking weights anyone had were a snapshot from March 2023, and roughly a third of the ranking heads had never been given a number at all. Then X published the live production values.
They are not a leak and not a snapshot. X's own README says it plainly:
The header of param.rs carries the sync stamp: last sync 2026-08-12T04:09:22Z.
One qualifier most coverage will drop. Every constant is declared as a feature-switch default, like param!(FavoriteWeight, f64, "rust_home_mixer_favorite_weight", 0.5);. X calls these the primary production values. A feature switch exists so some users can be served something else, so a reader inside an experiment bucket may have a different number applied to their feed. Read the file as the default configuration, captured on August 13.
The catch most of the internet is about to get wrong: every weight is a coefficient on a predicted probability. A report does not cost "minus 234 points." It contributes −234 × the model's predicted probability that a viewer reports the post. X's own code comment says the values are that large partly because the actions are rare.
How it works
- The engine is a faithful port of X's ranking_scorer.rs as of September 18, 2026: Σ(weight × P(action)) across 26 heads, then the negative-compression step, then author-diversity and out-of-network factors.
- The weights are X's production defaults, published August 13, 2026 and synced from X's own feature-switch configuration by cron. They are not a leak, and they are not the 2023 snapshot. An experiment bucket can be served an override, so treat them as the default configuration.
- Where the estimate enters. Phoenix is X's ranking model, and it predicts the probabilities. X ships phoenix/ in the repository, with the code that trains and runs the model plus synthetic data generation. What X does not ship is its trained model. Bangermeter estimates the probabilities instead: from real counts for likes, replies and reposts, and from content signals for everything else. The weights are X's. The probabilities are ours. The table above labels which is which, and so does every screen in the tool.
- Baseline probabilities are calibrated so that weight × P lands in a comparable band across heads, which is the balance X's own code comment implies. So the engine does not treat a 20.0 coefficient as forty times a 0.5 coefficient.
- Observed engagement rates get empirical-Bayes shrinkage, so a post with 1 reply in 19 views can't fake a viral score.
- Everything runs locally in your browser: zero network requests, no analytics, no account. MIT licensed, open source, 255 self-tests, X-DOM test suite in the repo.
We read the code so you don't have to trust screenshots
Bangermeter sits on four research passes, all documented in the repo's RESEARCH.md:
- Validation pass. Confirmed the 2023 weight table was the last sourced set X had released, and that the weighted-sum skeleton survived into X's 2026 open-source release.
- The folklore hunt. Every viral "leaked weight" was traced to its origin and debunked: "bookmark = 20× a like," "a link costs you 30 to 50 percent," "3+ hashtags = minus 40 percent," "block is minus 120, mute is minus 100." None survived source-tracing, and when the real values arrived on August 13 they matched none of it.
- The leak hunt. Swept court records, EU DSA proceedings, patents, scrubbed git history, and ex-employee disclosures. Found the two-snapshot weight history, the archived 2023 verified-author multipliers, a granted Twitter patent with a fitted engagement-vs-followers power law, and the fully open Community Notes algorithm with three causal studies. It found no third weight table, which is why folklore filled the gap.
- Same-day ingest, August 13, 2026. X published the production weights; Bangermeter shipped them the same day, with 112 tests asserting each value against param.rs line by line. The three earlier sweeps are what made that possible in hours instead of another year of guessing.
And the punchline that aged well: those viral Grok "algorithm reveal" screenshots were confabulations by construction. The weights were redacted from everything Grok could read. Now they aren't, and you can check every one yourself.
What moved since August 13
X changes this repository most weekdays, so every figure this page takes from X's code carries the date it was read. On August 25, 2026, X changed three weights in commit 0d3cdd8. There was no announcement, and the commit carries the same message as nearly every commit in that repository, so nothing in it said a weight had moved. The full account is in X changed three ranking weights and said nothing.
| Head | Before August 25 | As of September 18, 2026 |
|---|---|---|
| Video quality view | 0.05 | 0.0 |
| Dwell (binary) | 0.0 | 0.05 |
| Video open | 0.05 | 0.07 |
Two rules that both use ten seconds
X's reference code (phoenix/reference/world.py, which generates synthetic training data) marks a post dwelled, as of September 18, 2026, only when the viewer took an action on it and stayed at least 10 seconds. It marks not dwelled when the viewer took no action at all. A viewer who acts and leaves inside 10 seconds fires neither. Separately, MinVideoDurationMs is also 10,000, as of September 18, 2026. It measures the length of the clip itself, and watch time does not enter it. That head now carries a weight of 0.0, so the gate decides whether to add nothing. The two rules share a number and measure different things.
The small-account lane
EnableViewerColdStart ships on. On each For You request, at most one post gets its score raised to the 16th-highest score in the ranked list (index 15, counting from zero), if one qualifies: author at 1,000 followers or fewer, an original post (not a reply, not a repost), under 48 hours old, under 1,000 Home impressions, and already inside the top 85% of candidates with a nonzero score. These rules are as of September 18, 2026. The age limit rose from 24 to 48 hours on September 17. The lift sets a floor under the score. A post already scoring higher gains nothing, and later scorers can still move it. The pick happens on X's servers, so Bangermeter cannot tell whether your post was the one lifted, and it does not claim to.
Seven content inputs with no published weights
X now publishes what the ranker receives about a post's content, as of September 18, 2026: has_video, max_video_duration_ms, has_photo, media_count, weighted_text_len, newline_count, has_url. A URL counts as 23 characters whatever its length, and CJK characters and emoji count 2 (as of September 18, 2026). A media attachment's own t.co link is subtracted first, so posting an image does not make your post "have a link." newline_count is the first published evidence that post shape, as well as length, reaches the ranker. X publishes no coefficient for any of the seven, so Bangermeter lists them and does not score them.
Bookmarks: a model input with no weight
There is no bookmark weight. No coefficient multiplies a predicted bookmark, so a bookmark cannot turn into score the way a like or a reply can. The model still sees bookmarks: bookmark_count goes to Phoenix as a model input, in home-mixer/models/candidate.rs (as of September 18, 2026). The value model has no term to pay them with. "Bookmarks are 20× a like" is still folklore.
The panel moved out of the way
In version 0.10.2 the breakdown panel docks to the bottom-right corner at full height, so it no longer covers the post it explains. Click the badge of an open post to close the panel. Click a different post's badge to switch the panel to that post. Both work from the keyboard.
The trap: a big coefficient means the action is rare
Copy link is 20.0. A like is 0.5. Divide them and you get 40, and the sentence writes itself: a copy-link is worth forty likes.
I wrote that sentence, shipped it on this page, and my own red-team pass took it apart. Here is the correction, because it is the most useful thing on this page.
A weight is a coefficient. What a post earns from a head is weight × P(action), and X's code comment says the coefficient is partly a function of how rare the action is. Copying a link is far rarer than tapping like. Multiply each coefficient by its own probability and the order can flip.
Worked example, using Bangermeter's own estimated rates for a median post. On 10,000 impressions: about 50 likes and about half a copy-link.
Likes: 50 × 0.5 = 25.0
Copy links: 0.5 × 20.0 = 10.0
The like column wins by 2.5×. The probabilities in that example are ours, and they are estimates. The weights are X's.
"A copy-link is worth 40 likes" is the same sentence as "a reply is worth 27 likes," which is the folklore this release kills. Same arithmetic error, newer numbers.
The published weights tell you what X values per event. They do not tell you what to chase, because the probability half of the product lives in a model nobody outside X can run.
The largest single term on an ordinary post is not in the top of the table at all. NotDwelledWeight is −0.02 (as of September 18, 2026), near the bottom by coefficient, and it fires on most impressions you get. In X's reference code, which generates synthetic training data, a post leaves that term only when the viewer takes an action on it. A pause without an action still counts as not dwelled.
Receipts
Every figure in these cards is X's, as of September 18, 2026.
Who it's for
Creators and marketers
Who want to understand why posts perform instead of chasing folklore.
The algorithm-curious
Who will open "Show the math" and check the weight table themselves.
Anyone tired of the threads
The "I reverse-engineered the algorithm" ones. Now with a source you can check.
Get it
Tampermonkey userscript
Same tool, single file, for people who don't install extensions: extension/bangermeter.user.js in the repo.
Load it yourself
Clone the repo, open chrome://extensions, enable Developer mode, Load unpacked, pick the extension/ folder. Read the code before you run it. That's the point of it being open.
Score the next draft before it ships.
Free Chrome extension. MIT. Runs locally on x.com. Not affiliated with X Corp or xAI.
FAQ
Is a copy-link really worth forty likes?
Does a report really cost 234 points?
Is this the current algorithm?
Does it predict how many views I'll get?
What data do you collect?
Why do some huge accounts outperform their scores?
Can Bangermeter tell if my post got the small-account lift?
Is it free?
Bangermeter is an independent open-source project. Not affiliated with, endorsed by, or sponsored by X Corp or xAI. Scores are relative estimates: the weights are X's published production defaults, the probabilities they multiply are Bangermeter's own estimates. They are not predictions of reach.
Built by Ryan Lenk. More projects on the Side Projects page.