Bangermeter ⚡

Score any post with X's own published ranking weights.

A Chrome extension that scores posts on x.com using the real For You weights and the real scorer arithmetic from xai-org/x-algorithm — vm-ranker/params.rs (mirrored in home-mixer/params/param.rs) for the values, xai-value-model/scoring.rs for the math. Full source-traced findings: RESEARCH.md.

What you actually do with it

1. Scroll x.com. Every post picks up a ⚡ badge with two scores, 0–100 — C for what the content alone would be predicted to earn, E for what the post actually earned per view. 50 is a typical post: it is a comparison against a median post, not a grade, and not a prediction of reach.
2. Click the badge. You get the breakdown in plain English — which signals fired, what each reply and repost is worth, what age cost it — with every number cited to the line of X's published code it came from.
3. Start typing a post. A live meter scores your draft as you write, and + compare holds up to three variants so you can pick one before you send it.

Everything below is why you can trust the numbers.

Install

From the Chrome Web Store — Bangermeter. Installing opens a one-page quick start; the popup keeps a link back to it.

From source, to read it before you run it, or to work on it:

1. Open chrome://extensions
2. Enable Developer mode (top right)
3. Click Load unpacked → select the extension/ folder

August 13, 2026. X published the production weights for the For You timeline. Until that morning the only weights anyone had were a March/April 2023 snapshot, and about a third of the ranking heads had never been given a number at all. Bangermeter v0.9.0 replaced the entire weight layer with the published set. Everything below describes the new one.

What it looks like

Every post gets a badge. Click it for the breakdown: which signals fired, what each reply and repost is worth, the rescoring factors that applied, and a "Show the math" expander with the full weight × probability table.

What the scores mean

Every post gets up to two 0–100 scores (50 = median baseline post):

• C — Content score (prospective). What the ranker would be predisposed to predict from the content alone: baseline action probabilities per head, adjusted by detected content signals (question, thread, video, bare link, hashtag piles, engagement bait), run through the published weights and the real scorer arithmetic.
• E — Engagement score (retrospective). Applies the published weights to the post's actual rates (likes/reposts/replies ÷ views). Requires a visible view count.

Click the badge for the breakdown: per-head contributions (weight × P), detected signals with direction and provenance, rescoring factors, freshness, and a collapsible table of what the published weights actually say.

Creator tools (v0.10.0)

• Reply mode. Reply drafts are scored as replies — the ×0.75 in-network reply factor the compose meter previously missed — and the breakdown panel gets a reply-scoring section: out-of-network replies never reach For You at all (OONRetweetReplyFilter), replies are ineligible for the +15.0 mutual-follow boost, and replies to accounts over 250K followers (as of Sept 30, 2026 — X has raised that gate eleven times since August) are scored 0–3 by a Grok model whose rubric X withholds (the published inputs are listed; no invented direction — and no duration for the score-0 label, because none is published). Where a reply sorts inside a thread is disclosed as unpublished rather than guessed.
• Score history (opt-in). Off by default; enable it in the popup and opening a breakdown panel logs the scores to a local, capped list (chrome.storage.local, 200 entries — id, time, scores, an 80-char snippet). The popup shows the recent log with links back to each post, plus a Clear button. Nothing ever leaves the browser.
• Draft comparison. The compose meter gains a + compare button: save up to three variants of a draft (A/B/C), rewrite, and pick the best-scoring one. Variants live in memory only and vanish with the page.
• Posting-cadence warning. When an author has more than one post in the loaded stretch of feed, the panel reports the author-diversity attenuation production would apply — (1 − 0.25) × 0.5^k + 0.25: an author's 2nd post in the slate runs ×0.625, the 3rd ×0.44, to a ×0.25 floor — wherever in the slate they sit. Reported as context, not applied to the score (it is slate-relative and viewer-specific).
• Under the Hood import. Pilot-cohort users can import the JSON report X lets them download from x.com/i/jf/under_the_hood; the popup summarizes the visibility labels X itself applied (monthly aggregates — the report carries no post IDs). Parsed locally, stored locally, validated against the published label allowlist.

Methodology and honesty

• Formula: Σ(weight × P(action)) over the Phoenix head set, then offset_score — positive posts get +0.001, and any post whose weighted sum goes net-negative is rescaled into [0, 0.000897), which drops it below every positive-scoring post no matter what else it earned. Then the post-hoc factors: author diversity (1 − floor) × decay^k + floor, and the ×0.75 out-of-network factor. This is a direct port of X's value model (xai-value-model/scoring.rs, which replaced ranking_scorer.rs when X moved scoring into its vm-ranker service on Sept 24, 2026 — same math at published defaults), not an approximation of it.
• Weights: the published production set, following param.rs (stamp 2026-10-28T16:00:35Z). Likes 0.66 · replies 5.0 (20.0 on an original post from a mutual follow) · reposts 1.0 · quotes 5.0 · shares 2.0 · DM shares 5.0 · copy-link shares 20.0 · follow-author 4.0 · post clicks 0.3 · click dwell 0.4 · link opens 0.2 · photo expand / quoted click 0.05 · video open 0.02 · binary dwell 0.05 · dwell time 0.006 per second · post unexplored 0.015 · not-dwelled −0.02 · not-interested −52.272 · block −31.2 · mute −58.8 · report −234.0. Profile clicks, video-quality-view, quoted-vqv, react-with-video, and the continuation and profile-visit heads ship at 0.0.
• Five weights moved in param.rs between X's Oct 6 and Oct 13, 2026 syncs. Post unexplored went 0.02 → 0.015 (Oct 6 sync, published Oct 7). Video open went 0.07 → 0.02, dwell time 0.004 → 0.006 per second, and Likes 0.5 → 0.6 (Oct 12 sync, published Oct 13). Likes then went 0.6 → 0.66 and not-interested −47.52 → −52.272, another exact 10% (Oct 13 sync, published Oct 14). vm-ranker/params.rs still has the previous five numbers. The daily sync check treats that disagreement as drift and was not changed to ignore it. Four heads were added, all at 0. Two video-carousel weights also appeared in param.rs (1.0 and 0); they are not ranking heads, and the carousel is off by default.
• Three more weights moved, published 29 September 2026 from X's 28 September sync. Post clicks went 0.4 → 0.3, click dwell went 0.0 → 0.4, and not-interested went −43.2 → −47.52 — the first negative weight X has touched since publishing the table. Click dwell is named like a per-second term, but X's model config trains it as a yes/no head with a 10.0 threshold, so the 0.4 multiplies a probability: the chance a reader clicks in stays. That is always smaller than the chance of a click, so 0.4 against 0.3 does not mean staying is worth more than clicking. A browser cannot see clicks, so it is estimated.

Known limitations

• Count parsing prefers a locale word table (16 locales today; more join only with sourced strings) and falls back to the locale-independent data-testid buttons, so counts and views survive even on locales the table does not know.
• Reply detection uses whichever signal the surface actually provides: the "Replying to" label on timelines, and position in the thread inside a conversation or on with_replies, where X renders no label at all.
• Engagement-bait detection covers English, Hinglish and Devanagari imperative calls to action. The Devanagari patterns are not transliterations of the Hinglish ones — Hindi puts the call to action last, so "like karo agar X" is an English calque that barely occurs, and the real form is "…तो लाइक करें" or "कमेंट बॉक्स में जरूर". That anchor does the work: on X the bare verb phrase is usually an argument ("पहले पढ़ो, फिर कमेंट करो" — first read it, then comment) rather than a request for engagement, so bare कमेंट कर / फॉलो कर are deliberately excluded. Validated against real Hindi posts collected from X: zero false positives, conservative recall.
• The rhetorical-question genre ("Kya …?", "X ya Y?") is deliberately not detected. Tested against 228 real posts, the structural rule for it flagged exactly one — a top-quartile post — and missed the actual bait. ALL-CAPS detection is Latin-only by nature, since caseless scripts have no capitals to count. Reply detection on timeline posts still needs the localized "Replying to" marker; reply drafts are detected structurally (dialog order, status-page URL), which is locale-independent.
• The E score needs a visible view count (hidden on some surfaces).
• Viewer-specific factors can't be observed: in-network status and mutual-follow status. Out-of-network and mutual-follow are popup toggles instead.
• Quote counts aren't exposed in the timeline DOM, so the quote head (5.0) is estimated rather than measured.
• X changes its DOM without notice; selectors have documented fallbacks.

Project files

File	Purpose
extension/	The Chrome extension (MV3, vanilla JS, no build step)
extension/weights.js	Single source of truth: weight layer + estimator layer, all provenance-tagged
extension/scoring.js	Pure scoring engine (direct port of X's value model, xai-value-model/scoring.rs)
extension/content.js	Badges, breakdown panel, compose meter
extension/background.js	Service worker. One listener: open the quick start on first install, never on update
extension/welcome.html	The quick start itself — self-contained, loads nothing over the network
extension/test.html	Engine self-test — open in any browser (364 assertions)
extension/fixture.html	X-DOM fixture harness for the content script
extension/fixture-thread.html	Reply-detection harness — asserts the conversation, with_replies and home-timeline surfaces separately, because X marks a reply differently on each. Needs serve-fixtures.js (it reads location.pathname)
extension/serve-fixtures.js	Tiny static server for the harnesses, including the x.com-shaped paths the reply-detection cases need
extension/fixture-compose.html	Compose-meter visibility harness — asserts the draft meter survives X's scrolling, overflow-hidden compose dialog (it prints its own pass count)
extension/bangermeter.user.js	Single-file Tampermonkey build (generated — see store-assets/make-userscript.ps1)
packages/bangermeter-rank/	Ranking modes (account audit, C checklist, E retrospective, stealability), score cards pinned to the param.rs sync, the daily weight-sync check, the customer audit report, and the copy and Numbers gates. Zero dependencies. See its README
skills/audience-readout/	Claude skill: collect one account's real posts, score them, and write a read-out. Carries the analytical rules (rates not counts, confound checks) and the privacy rule — never commit an archive. Copy to ~/.claude/skills/ to install — see its SKILL.md

Nothing numeric is typed into the store art. make-promos.ps1 and make-screenshot.ps1 read their weights and ratios from extension/weights.js through weights-export.js, so a tile cannot keep publishing a value after the code has moved on — which is exactly what happened when those numbers were hardcoded.

Verification status

• Engine math: 364/364 self-tests pass (test.html). Every published weight and its feature-switch parameter name is asserted individually against param.rs, or vm-ranker/params.rs for the three heads only it publishes, so a silent transcription error fails the suite rather than shipping. param.rs moved five heads between the Oct 6 and Oct 13, 2026 syncs (published Oct 7–14), and vm-ranker/params.rs did not, so the two upstream files disagree on those five at the 2026-10-16T16:00:32Z sync; weights.js follows param.rs. Four heads were added (29 in the roster), plus two video-carousel weights that are not heads. X deleted a 26th head, cont_active_secs_5m_residual_norm (weighted 0.0), in its Sept 23 sync, published Sept 24. packages/bangermeter-rank runs sync daily in CI to catch the next change.
• Locale strings (reply markers and count words for 16 locales) are transcribed from X's own production i18n bundles (abs.twimg.com/responsive-web/client-web/i18n/*), fetched Aug 25, 2026 and cross-checked against Wayback captures of the same bundles. Locales join the table only with a sourced string — the same rule the weight layer follows. The "Visibility limited" label (message key b3296688, all 16 locales) was added Sept 30, 2026 from the bundles named in the Sept 12 Wayback capture of the x.com/i/flow/login chunk manifest. The live logged-out page now serves X's x-web client, which carries no such manifest.
• Adversarially reviewed at v0.9.0 by three independent passes — weight transcription, Rust-to-JS arithmetic fidelity, and a stale-claim sweep. The arithmetic pass swept ~654k generated inputs for NaN, out-of-range and non-monotonic scores and found none, and used mutation testing to prove which assertions were load-bearing. Findings from all three are fixed in this release; the notable ones are recorded in RESEARCH.md.
• Content script (badges, panel, meter, virtualized lists, quote-tweet scoping): exercised against the DOM fixture (fixture.html). The fixture is an eyeball harness — it renders and must not throw; it carries no assertions.
• Live x.com: verified in a logged-in session, Aug 2026. Badge, panel and per-head copy render correctly against the live DOM, and a spot-checked post matched the engine's arithmetic to the rounded point. Live use is also what surfaced the two compose-meter defects the fixture had not modeled. Selectors target article[data-testid="tweet"], the action-bar role="group" aria-label, and tweetTextarea_*, each with fallbacks.

Credits

• First-run onboarding — @harvey_control, who installed Bangermeter, kept using it to write replies, and then took the trouble to say the thing a happy user almost never bothers to say: that he could not work out quickly what it was for. That is more useful than a bug report, because nobody files a ticket for confusion — they just leave. He was right about the cause, too. This README and the store listing were ordered provenance-first, every number sourced before you learn what to do with it, which reads as rigor to someone already convinced and as a wall to everyone else. Installing compounded it by completing in silence. Both changed in 0.10.1: the quick start opens itself, and the first thing you read here is what the thing does rather than where the weights came from. A tool nobody can get into is not a rigorous tool, just an unused one.

• Hindi and Hinglish guidance — Shivangi Bhattacharjee, who worked through the bait phrasing with us. The useful finding was structural rather than lexical: Hindi puts the call to action at the end of the sentence, so the Hinglish "like karo agar…" order is an English calque and a Devanagari pattern built by transliterating it would never have matched anything. Every Hindi pattern the extension ships anchors on तो / जरूर instead, and the bare verb is deliberately left alone because on X it usually belongs to an argument rather than a request for engagement.