45 facts from X's public ranking code

In checked-in source, the compile-time value is 50: TopKScoreSelector keeps at most this many scored candidates per feed request.

code fact · compile-time constant

In checked-in source, the compile-time constant RESULT_SIZE is 35. It is the Phoenix scored-posts pipeline's result_size (phoenix_candidate_pipeline.rs:1016), and the ads blenders (multi_risk, partition_organic, time_gap, safe_gap) cut the blended post-and-ad list to it, so there the 35 includes ads.

code fact · compile-time constant

In checked-in source, the For You pipeline's result_size() cap per response is 47 items: RESULT_SIZE 35 + FEED_MODULE_SLOTS 4 + MAX_JETFUEL_FRAMES_PER_RESPONSE 8, so non-post module and frame slots count toward it.

code fact · compile-time constant

In checked-in source, MAX_POST_AGE is 48 * 60 * 60 = 172800 seconds (48 hours); the Phoenix candidate pipeline's AgeFilter uses it as the maximum post-candidate age, measured from the tweet ID timestamp.

code fact · compile-time constant

At this tree VMRanker::enable (vm_ranker.rs:80-82) returns the EnableRanking feature switch (rust_home_mixer_enable_ranking, checked-in default true, home-mixer/params/param.rs line 257), its only reader here. At a389166f6cf5da70a286b568c87695d4dcdce3a1 VMRanker read EnableVMRanker (default true) and EnableRanking gated ranking_scorer.rs; EnableVMRanker is absent here.

code fact · feature switch default

The checked-in feature-switch default for the per-viewer served-history exclusion window is 10 minutes (rust_home_mixer_exclude_served_tweet_ids_duration_minutes).

code fact · feature switch default

The checked-in feature-switch default for ExcludeServedTweetIdsNumber, the cap on how many recently served post IDs from a viewer's own served history are excluded from that viewer's candidates, is 100.

code fact · feature switch default

In checked-in source the compile-time constant FOLLOWING_MAX_RESULT_SIZE is 110 (100 at a389166f6cf5da70a286b568c87695d4dcdce3a1), so FOLLOWING_POST_FETCH_SIZE is 101 (110 minus FOLLOWING_ADS_TOP_K 9) and FOLLOWING_PIPELINE_RESULT_SIZE is 112; the separate RANKED_FOLLOWING_MAX_RESULT_SIZE stays 38.

code fact · compile-time constant

In checked-in source, home-mixer's weights_for hard-codes enable_author_diversity false, author_diversity_decay 1.0, author_diversity_floor 1.0 and oon_rescore_in_network_replies_retweets false (lines 40-43), and its weighted_score calls only fuse_heads (line 83), never post_fusion_multipliers, so the local score applies no repeated-author decay or out-of-network factor.

code fact · compile-time constant

In checked-in source, the vm-ranker binary's --dpp-enabled startup flag defaults to false; without it no DPP (diversity reranking) context is built, and the per-request DppEnabled switch cannot create one.

code fact · startup flag default

vm-ranker's checked-in feature-switch defaults are EnableAuthorDiversity true, AuthorDiversityDecay 0.5 and AuthorDiversityFloor 0.25; within one request the code's formula (1-floor)*decay^k+floor gives an author's 1st to 4th highest-scored candidates multipliers 1.0, 0.625, 0.4375 and 0.34375. Params resolve only when startup flag config_sync_enabled (default false) is on.

code fact · feature switch default

vm-ranker's checked-in feature-switch defaults are OonWeightFactor 0.75, TopicOonWeightFactor 0.5 (topic requests) and EnableOonRescoreForInNetworkRepliesRetweets true; oon_applies (xai-value-model/scoring.rs line 166) marks candidates out-of-network for the viewer and, while that switch is on, in-network replies and reposts, leaving unknown status unmarked. Params need config_sync_enabled.

code fact · feature switch default

In this checked-in code, a viewer's cached feed candidate list counts as usable (has_cached_posts) only if it holds at least 500 cached post candidates (const MIN_CACHED_POSTS_THRESHOLD = 500).

code fact · compile-time constant

In checked-in code, the cached-posts query hydrator stops waiting for its Redis GET after 300 ms (a compile-time constant); the hydrator runs only when the EnableCachedPosts param is on, whose checked-in default is true.

code fact · compile-time constant

In checked-in source, the side effect that caches a viewer's already-scored feed candidates in Redis sets a 180-second expiry (const REDIS_TTL_SECONDS passed to set_ex); it runs only when is_prod() is true and the query had no cached posts.

code fact · compile-time constant

Code timeout is 200 ms; timeout or client error uses the external ViewerData::default(), while an explicit in_network_only request is OR-preserved.

code fact · compile-time constant

In checked-in source, the feature switch under_the_hood_report_enabled defaults to false.

code fact · feature switch default

Checked-in feature-switch fallback defaults: under_the_hood_report_takedowns_enabled false (line 22), under_the_hood_report_max_month_bucket 0 (line 19), and the report's master switch under_the_hood_report_enabled false. Legal-withholding and copyright/DMCA takedown rows are added only when the takedowns switch is on (underTheHoodReport.User.strato:446); a bucket of 0 or less sets no month cap.

code fact · feature switch default

Checked-in report JSON needs a non-suspended account strictly older than 365 days with at least 10 eligible posts (lines 18-19) in one prior UTC month: last month once 10 days pass from 00:00 UTC on its last day (line 20), else the month before. max_month_bucket (default 0, no cap) can move it earlier; incomplete data falls back one month, and still-incomplete data returns nothing.

code fact · compile-time constant

The checked-in daily post takedown labels job keeps a post's takedown observations (PCTD and DMCA audit deltas plus tweet events) only if the post was created within the window before the as-of day and the event precedes the sooner of creation plus the window or day end (lines 208-212). The window is job argument postObservationDays, 7 when not passed (line 244); rows go to the authoring day.

code fact · startup flag default

postLabelRow and accountLabelRow emit rows only for allowlisted labels, plus takedown reasons only when the under_the_hood_report_takedowns_enabled feature switch (checked-in default false) is on (lines 242-312), so this tree does not establish that an empty report means a creator's distribution was unaffected.

derived from the code · feature switch default

Upstream's line-1 comment reads 'mirrored from config feature-switch defaults; last sync 2026-09-24T16:24:49Z' (2026-08-12T04:09:22Z at a389166f6cf5da70a286b568c87695d4dcdce3a1). README lines 408-410 say experiments run on a small share of timeline traffic, many values come from a config system, and cron scripts set repo defaults to primary production values.

code fact · feature switch default

apply(score, weight) returns score.unwrap_or(0.0) * weight (line 57); compute_weighted_score (line 65) sums 25 such terms, each a weight times one Phoenix predicted head in candidate.phoenix_scores, from favorite_score through report_score to post_unexplored_score. No raw engagement-count field is an operand of this fusion step; weight values come from request params.

code fact · compile-time constant

At the checked SHA the feature-switch default BidirectionalFollowReplyWeightBoost is 15.0; ValueModelWeights::reply_weight_for (xai-value-model/weights.rs line 78) adds it to ReplyWeight (default 5.0) only for non-reply, non-repost candidates whose author and viewer follow each other, as tagged when EnableBidirectionalFollowHydration is on and posts are not cached. The dwell boost defaults to 0.0.

code fact · feature switch default

An upstream comment (param.rs lines 270-276, repeated at xai-value-model/scoring.rs lines 32-38; README lines 39 and 349) says each weight multiplies the predicted probability of an action, or a predicted value such as watch time, not raw engagement counts, and that reading weight ratios as count equivalences is incorrect. The scoring code is consistent with it.

code fact · live state unknown

Checked-in feature-switch defaults (synced 2026-09-24) set ReportWeight -234.0 (line 401) and FavoriteWeight 0.5 (line 295). The report weight is several hundred times larger, but this tree does not establish any exchange rate between report and like counts: both multiply a viewer's predicted probability, and lines 273-280 reject the count reading and cite a far lower report base rate.

derived from the code · feature switch default

An upstream comment (param.rs 282-294, repeated at xai-value-model/scoring.rs 44-56) calls it a misunderstanding that coordinated blocking or reporting by bad actors will significantly lower reach: predictions are per viewer, so such reports mainly affect viewers similar to the reporters, and only actions on Home-served posts count. Upstream's reasons; this tree does not enforce the Home rule.

code fact · live state unknown

Checked-in feature-switch defaults are EnableViewerColdStart true (line 491) and EnableColdStartThompsonSampling false (line 497). Then AuthorColdStart picks, per request, the eligible candidate with the highest current score and sets it to max(existing, target), never lower and possibly unchanged (author_cold_start.rs 315-335); switched off, scores return unchanged.

code fact · feature switch default

cold_start_base_eligible (line 171) is a necessary pre-filter for the cold-start lift: it admits only non-reply, non-repost posts whose author follower count is known and at or below ColdStartFollowerCap (feature-switch default 1000, param.rs line 471); it does not exclude quotes. Admitted posts still face the other apply_cold_start gates before at most one per request is lifted.

code fact · feature switch default

In every viewer arm, apply_cold_start admits a candidate only if post age is at most ColdStartMaxPostAgeSecs (feature-switch default 172800 s; 86400 and treatment-arm only at a389166f6cf5da70a286b568c87695d4dcdce3a1), a known view_count_on_home is below ColdStartImpressionThreshold (default 1000), and its position is below floor(0.85 x non-zero scores); base, corpus and enable gates also apply.

code fact · feature switch default

With feature-switch defaults ColdStartSlotMin 15 and ColdStartSlotMax 16 (param.rs 459-468), cold_start_target picks zero-based rank 15, the 16th-highest pre-policy score, with no target under 16 candidates. The lift is max(existing, target), a score floor, not a placement, and the value model repeats it on its own scores, so this tree does not establish a reserved feed slot for small accounts.

derived from the code · feature switch default

Brazil2026ElectionFilter has no enable override (trait default true) and a compiled-in list of 2786 account ids its header ties to 2026 Brazilian Electoral Court reports. In PhoenixCandidatePipeline (line 381), it removes a candidate when a listed account the viewer does not follow is its author, reposted, quoted or ancestor author; it checks no viewer country and reads no feature switch.

code fact · compile-time constant

In checked-in source FavHoldoutFilter (phoenix pipeline) runs only when EnableFavHoldout is on (feature-switch default false, param.rs line 822); it then removes a candidate when a deterministic hash of original post id and viewer id mod 100 is below a like-count percentage: 0 at one like or fewer, 2 at two, stepping to 14 at exactly 3000 and 15 above. It has no follow exemption.

code fact · feature switch default

For one tweet id with no high- or medium-risk label, compute_verdict returns MediumRisk if it has neither GROK_SFA nor GROK_NSFA_LIMITED (lines 57-61), or its id is at or above PTOS_CUTOFF_TWEET_ID (line 43, about 2026-05-12 19:00 UTC) without PTOS_REVIEWED. After top-K selection the hydrator takes the worst verdict over post, quote and ancestors; ad and response code read it, no scorer.

code fact · compile-time constant

In the checked-in Rust visibility-filtering service, proto_action maps every verdict to Drop, Interstitial or Allow (tombstone to Drop, engagement-limit-only to Allow), and ActionSpec (rules/rule_spec.rs line 112) has no downrank or score-multiplier variant. This covers only its output: the VF client and legacy reference still carry Downrank and Avoid, which home-mixer's vf_filter.rs (line 24) keeps and no home-mixer scorer reads.

code fact · compile-time constant

In checked-in source, TIMELINE_HOME_RECOMMENDATIONS_POLICY is the shared Home rules plus nine recommendation-only groups (registry.rs 140-156), one dropping posts labelled DoNotAmplify, MaliciousUrl or SpamHighRecall for all but the author (tweet_rules.rs 233-252). home-mixer requests it for out-of-network candidates and every candidate's ancestors and quotes; in-network ones get TimelineHome.

code fact · request-time condition

In checked-in rules, OON_USER_LABEL_DROPS (lines 127-176) sits only in TimelineHomeRecommendations and ImmersiveExpandedRecommendations, not TimelineHome. By author label, not post content, it drops posts of authors labelled SpamHighRecall, Compromised, ReadOnly or ImpersonationHighPrecision for all but the author, and AbusiveHighRecall or DoNotAmplify for non-followers, logged-out included.

code fact · compile-time constant

In checked-in home-mixer, should_drop_ancillary (lines 140-170) flags a candidate when this viewer's VF result for any non-tombstoned ancestor, its quoted post or reposted source is Drop, Tombstone or NotEvaluated; AncillaryVFFilter then removes it from that viewer's For You and Following results. A failed or missing lookup, or Allow, Interstitial, Avoid or Downrank, sets no flag.

code fact · compile-time constant

In checked-in code DropStaleTweetsRule drops a non-repost post whose edit-control data lists a later id as the chain's last entry, for Audience::Everyone (author not exempt) with UnspecifiedReason (tweet_rules.rs 554-562; models/tweet.rs 42-52; no edit data means not superseded). It is in TIMELINE_HOME_SHARED_RULES (rules/registry.rs line 132), used by TimelineHome and TimelineHomeRecommendations.

code fact · compile-time constant

In checked-in policies with TAKEDOWN_DROPS (Home, Home recommendations, hydration, immersive), a post drops for every non-author viewer when a LegalRequest (not xy) or UnspecifiedReason code equals the request's country_code or is worldwide (xx, xy; DMCA counts as xy), or a BystanderReport code equals that country (models/tweet.rs 62-110). With no request country, only worldwide codes match.

code fact · request-time condition

For non-evergreen sources, executeOp skips community posts, replies, reposts and VF-dropped or adult posts (lines 446-459) before any 1fav, 32fav or post-creation index event; replies still reach the earlier-built search_unfiltered index. A like on an eligible original calls build1FavIndex (line 463), with no like-count gate (a389166f6cf5da70a286b568c87695d4dcdce3a1 needed one).

code fact · compile-time constant

Checked-in _home_direct_packed_base (line 244) sets enable_is_author_followed_by_viewer, enable_is_author_following_viewer and enable_engagement_counts True (290-292; dataclass default False for counts). Feature prep adds each follow flag as a learned embedding and turns like, reply, repost, quote and view counts of history and candidate posts into log2(count + 1) / 32 inputs to a 64-unit MLP.

code fact · checked-in default

In checked-in code as_tweet_info (line 295) sets content_features on each candidate's Phoenix TweetInfo (line 352) and, for quote candidates, quoted_content_features (line 353): has_video, max_video_duration_ms, has_photo, media_count, weighted_text_len, newline_count and has_url (a URL-count heuristic). Under phoenix/ only the proto mentions them, and the served model is not in this tree.

code fact · compile-time constant

GROK_GEMMA_FOLLOWER_SPLIT is 250_000 (inline 15000 at a389166f6cf5da70a286b568c87695d4dcdce3a1): replies whose parent and root authors both have at most 250,000 followers are eligible for the Gemma spam plan, others for Grok reply ranking. Both plans end in one write task: a first or lowering 0.0 score applies RiskyHighVizReply unless the author is high-page-rank, grey-badge or a test user.

code fact · compile-time constant

TaskMultiStepReplySpamFilter only gates eligibility: a reply 2+ levels deep whose parent is by the same replier, under another root author with 1000+ followers (line 26), excluding high-page-rank and grey-badge repliers, goes to an LLM judge (prompt not in the repo). If the newest reply and 2+ of the replier's posts are flagged, they get a 0.0 score and, test users aside, RiskyHighVizReply.

code fact · compile-time constant