Unofficial, community-maintained project. Not affiliated with, endorsed by, or operated by X Corp. or xAI. Notes are generated from the public git history of xai-org/x-algorithm and may contain mistakes; the upstream diff is the source of truth.
Plain-English developer notes for every sync of xai-org/x-algorithm, the open-source X “For You” recommendation algorithm. Upstream commits are published by a CI agent and all carry the same message (“Open-source X Recommendation Algorithm”), so this site reads each diff and writes down what actually changed: new or removed filters, scorers and candidate sources, ranking-weight and parameter changes with old → new values, and new feature flags.
45 entries · newest first · subscribe via Atom
Defaults of *Weight param!s at upstream HEAD (home-mixer/params/param.rs and vm-ranker/params.rs). These multiply the model's
predicted probabilities of each action for a given viewer, not raw engagement counts (see the upstream README). Production values can be overridden by feature switches.
| Weight | home-mixer | vm-ranker | Last changed |
|---|---|---|---|
BidirectionalFollowDwellWeightBoost | 0.0 | 0.0 | — |
BidirectionalFollowReplyWeightBoost | 15.0 | 15.0 | — |
BlockAuthorWeight | -31.2 | -31.2 | — |
ClickWeight | 0.3 | 0.3 | Sep 29, 2026 (was 0.4) |
ContClickDwellTimeWeight | 0.4 | 0.4 | Sep 29, 2026 (was 0.0) |
ContDwellTimeWeight | 0.004 | 0.004 | — |
DwellWeight | 0.05 | 0.05 | Aug 25, 2026 (was 0.0) |
FavoriteWeight | 0.5 | 0.5 | — |
FollowAuthorWeight | 4.0 | 4.0 | — |
MuteAuthorWeight | -58.8 | -58.8 | — |
NotDwelledWeight | -0.02 | -0.02 | — |
NotInterestedWeight | -47.52 | -47.52 | Sep 29, 2026 (was -43.2) |
OpenLinkWeight | 0.2 | 0.2 | — |
PhotoExpandWeight | 0.05 | 0.05 | — |
PostUnexploredWeight | 0.015 | 0.02 | Oct 7, 2026 (was 0.02) |
ProfileClickWeight | 0.0 | 0.0 | — |
ProfileVisitSecsWeight | — | 0.0 | — |
QuoteWeight | 5.0 | 5.0 | — |
QuotedClickWeight | 0.05 | 0.05 | — |
QuotedVqvWeight | 0.0 | 0.0 | — |
ReplyWeight | 5.0 | 5.0 | — |
ReportWeight | -234.0 | -234.0 | — |
RetweetWeight | 1.0 | 1.0 | — |
ShareViaCopyLinkWeight | 20.0 | 20.0 | — |
ShareViaDmWeight | 5.0 | 5.0 | — |
ShareWeight | 2.0 | 2.0 | — |
UserVideoContinuationWeight | — | 0.0 | — |
VideoContinuationWeight | — | 0.0 | — |
VideoOpenWeight | 0.07 | 0.07 | Aug 25, 2026 (was 0.05) |
VqvWeight | 0.0 | 0.0 | Aug 25, 2026 (was 0.05) |
bs_per_device 256→512 (xrecsys.py) and ep 256→128 (xrecsys.py).PredictNextActionsRequest.feed_cursor_direction and PredictNextActionsRequest.feed_request_context).phoenix/ (+1,215/−217), grox/ (+12/−0) and home-mixer/ (+1/−1).Full analysis → Summary: automatic analysis
EnableVideoCarousel (default off), plus VideoCarouselPosition = 6 and VideoCarouselFatigueMinutes = 60.VIDEO_CAROUSEL_EXTRA_CANDIDATES = 30 extra video candidates (top by predicted video-open). The new VideoCarouselFilter and util/video_carousel.rs then pick 3–5 non-reply videos whose aspect ratio is below 1.0 (the code's is_vertical check), ordered by predicted video-open probability. No carousel is shown if fewer than 3 qualify.video_aspect_ratio_hydrator supplies the aspect ratio. FEED_MODULE_SLOTS 4 → 5.ads/drops.rs. Large Phoenix training/model changes in the same sync.Full analysis → Summary: hand-written notes, checked against the diff
EnablePopularPostsSource now on by default. The README adds it as a third out-of-network source next to Phoenix retrieval and SimClusters: "a shared list of popular recent posts, initially from the most-followed active accounts."PostUnexploredWeight 0.02 → 0.015.Full analysis → Summary: hand-written notes, checked against the diff
0.0 for now: VideoContinuationWeight, UserVideoContinuationWeight and ProfileVisitSecsWeight, defined in vm-ranker/params.rs. The shared xai-value-model crate gets matching heads (its weight list grows from 25 to 28 entries). Home-mixer's own value_model.rs hard-codes the two video-continuation weights to 0.0, so for now these heads don't change scores.PostUnexploredIncludeOutOfNetwork (vm-ranker/params.rs, default off). Without it, the value model's post-unexplored bonus applies only to in-network candidates.PopularPostsSource now reads from a precomputed PopularPostsCache (new util/popular_posts.rs, filled by new popular_posts_job.rs / popular_authors_job.rs) instead of querying Thunder on each request. PopularPostsTopAuthors, PopularPostsMaxPerAuthor and PopularPostsMaxRepliesRepostsPerAuthor leave param.rs.EnableUnderTheHood and EnableHomeExcursionScores (both off).RetrievalCandidatesKafkaMaxCandidates 200 → 100000. VMRankerClusterId "Experiment6" → "Experiment1".Full analysis → Summary: hand-written notes, checked against the diff
PopularPostsSource (home-mixer/sources/popular_posts_source.rs), behind EnablePopularPostsSource (default off). It fetches recent posts by popular authors (PopularPostsTopAuthors = 1000) from Thunder, capped at PopularPostsMaxResults = 500 and PopularPostsMaxPerAuthor = 3, with no replies or reposts (PopularPostsMaxRepliesRepostsPerAuthor = 0).EnableAdsBrandSafetyVerdictV2 goes back to on (it was switched off on Sep 16).EnablePhoenixOonReplies, AuthorExplorationBonus = 0.0, EnableAdsAuthorBrandSafetyFallback, VMRankerSendPacingInputs.phoenix_reply_ancestors_hydrator and a served_event_kafka side effect.Full analysis → Summary: hand-written notes, checked against the diff
SidSource (home-mixer/sources/sid_source.rs) that retrieves posts through a new SidRetrievalClient (seeded with up to SidSourceMaxSeeds post IDs taken from the viewer's signals via post_signal_ids). It is gated by EnableSidSource (default off) with SidSourceMaxResults = 800, SidSourceMaxSeeds = 50, SidSourceMaxPerSeed = 100 and SidSourceMinPrefixDepth = 3.TestBoost and TestCap params.Full analysis → Summary: hand-written notes, checked against the diff
FeedSurveyFatigueMinutes = 1440.FeedSurveyFatigueHours.(2, failed()), 1→2 (fallback_cache.rs), (1, failed()), 2→1 (fallback_cache.rs) and fn verdict_cases() -> [(Verdict, Projected); 15] { 14→15 (treatment.rs).visibility-filtering/ (+1,566/−300), abuse-enforcement-service/ (+713/−150), phoenix-rankall/ (+139/−33), phoenix/ (+45/−11) and home-mixer/ (+9/−11).Full analysis → Summary: automatic analysis
This sync changes the defaults of the author cold-start mechanism (home-mixer/scorers/author_cold_start.rs). Per request, it picks one eligible fresh post from a smaller account and raises its score to whatever score sits at a fixed rank, which is index 15 given ColdStartSlotMin = 15 and ColdStartSlotMax = 16.
ColdStartFollowerCap 1000 → 50000: authors with up to 50k followers are now eligible.ColdStartImpressionThreshold 1000 → 200: the post must have fewer than 200 Home impressions.ColdStartMaxPostAgeSecs 172800 → 7200: the post must be under 2 hours old (was 48h).LowImpressionsMaxPositionRatio 0.85 → 0.97: candidates from deeper in the ranked list can qualify.EnableColdStartThompsonSampling now on by default. Instead of taking the highest-scored eligible post, the scorer samples a Beta(0.75 + likes, 49.25 + impressions − likes) reward per post, keeps the top ColdStartTsTopK (2), and boosts the higher-scored of those.PhoenixColdStartMaxResults 0 → 200: the Phoenix cold-start retrieval path is switched on.Full analysis → Summary: hand-written notes, checked against the diff
home-mixer/params/param.rs and vm-ranker/params.rs: ClickWeight 0.4 → 0.3. ContClickDwellTimeWeight 0.0 → 0.4, so predicted time spent after clicking into a post now counts. NotInterestedWeight -43.2 → -47.52, a 10% larger penalty.author_cold_start.rs the MoE "ranking policy" that zeroed arm-gated retrieval candidates for non-treatment viewers is removed, along with the author_policy_zeroed field. Holdout viewers no longer exclude arm-gated retrieval candidates from cold-start eligibility.PhoenixRetrievalExcludeSeenPosts (default off). When on, Phoenix retrieval is sent the viewer's seen/served post IDs and impression bloom filters so it can skip posts already shown.client_seen_post_ids).UseEngagementCounterViewCountForImpressionBoost.abuse-ledger-service/. When an enforcement is overturned on appeal, it is not re-applied for a fixed period.Full analysis → Summary: hand-written notes, checked against the diff
safety_label_hydrator and tes_hydrator.gizmoduck_hydrator → author and viewer_hydrator → viewer.GEMMA_REPLY_SPAM_MIN_FOLLOWERS 150_000 → 125_000 (constants.py) and FOLLOWER_COUNT_THRESHOLD_FOR_SPAM_DETECTION 100_000 → 125_000 (task_filter.py).rows 14→15 (execute.rs) and timeline_home_hydration 10→11 (plan.rs).PostActionScores.action_indices, PostActionScores.found, PostActionScores.num_posts, PostActionScores.post_action_masks and 14 more).abuse-enforcement-service/ (+6,060/−7), visibility-filtering/ (+2,177/−1,937), phoenix/ (+811/−689), home-mixer/ (+62/−8) and grox/ (+20/−11).Full analysis → Summary: automatic analysis
conversation_control_hydrator, exclusive_content_hydrator and socialgraph_hydrator.KEY_FRAMES_UNCONDITIONAL_FAV 512 → 1024 (constants.py).TIMELINE_HOME_RECOMMENDATION_ONLY_RULES 5→9 (registry.rs).BRAZIL_2026_ELECTION_USER_IDS (brazil_2026_election_filter.rs) grew from 2,776 to 2,786 entries (per its in-file test assertion).AdInfo.unified_card_type).visibility-filtering/ (+6,329/−5,545), grox/ (+534/−153), under-the-hood/ (+616/−0), phoenix/ (+256/−52) and home-mixer/ (+23/−3).Full analysis → Summary: automatic analysis
xai-value-model cratehome-mixer/scorers/ranking_scorer.rs (1,255 lines) is deleted. It's replaced by a 114-line home-mixer/scorers/value_model.rs that maps feature-switch params into a ValueModelWeights struct and calls fuse_heads from a new top-level crate, xai-value-model/ (scoring.rs, weights.rs, inputs.rs, …).weights_for hard-codes enable_author_diversity: false, oon_rescore_in_network_replies_retweets: false and multiplier_pre_offset: false.home-mixer/params/param.rs. Eleven of them (author diversity, the OON weight factors, MultiplierPreOffset, WeightPerturbation*, …) are still defined in vm-ranker/params.rs. The other eleven aren't defined as params anywhere else, including EnableVMRanker, CachedPostsReuseWeightedScore, EnableMultiplicativePostUnexplored and the VMRankerDpp* settings.VMRankerClusterId changes from "Experiment3" to "Experiment6".Full analysis → Summary: hand-written notes, checked against the diff
vm-ranker/ gets its own params.rs with 44 param! defaults. These include 27 ranking weights that mirror home-mixer's names, keys and current defaults (all of home-mixer's weights except ContActiveSecs5mResidualNormWeight; e.g. FavoriteWeight = 0.5, ReplyWeight = 5.0), plus DPP settings (DppEnabled on, DppTheta = 0.65, DppMaxSelectedRank = 150) and author-diversity settings (EnableAuthorDiversity on, AuthorDiversityDecay = 0.5, AuthorDiversityFloor = 0.25).vm-ranker/scoring/value_model.rs builds value-model weights from those params (weights_from_params). It is gated by ComputeValueModel, which defaults to off. New ranking_config.rs and config_sync.rs files are added alongside.vm_ranker_debias_payload.rs and a VMRankerSendDebiasInputs flag (off). PhoenixMOEMaxResults 200 → 0 and PhoenixMoeColdStartMaxResults 0 → 200. PhoenixRetrievalMOEInferenceClusterId "Experiment3Memy04" → "Experiment2Memy04".GROK_GEMMA_FOLLOWER_SPLIT 200_000 → 225_000.Full analysis → Summary: hand-written notes, checked against the diff
VMRankerComputeValueModel (default off) and VMRankerSendValueModelInputs (default off).NewUserOonWeightFactor = 0.00001, PhoenixMoeColdStartMaxResults = 0, RetrievalCandidatesKafkaMaxCandidates = 200, RetrievalCandidatesKafkaSamplePercent = 5.0 and SimclustersMaxCandidateAgeHours = 48.vm_ranker_request.retrieval_candidates_kafka_side_effect.FOLLOWER_COUNT_THRESHOLD_FOR_SPAM_DETECTION 50000 → 80_000 (task_filter.py) and HYDRATION_TIMEOUT Duration::from_millis(390) → Duration::from_secs(1) (mod.rs).(4, tonic::Code::Unimplemented), 82→4 (evaluate_tweets.rs) and Some(&2) 1→2 (metrics.rs).visibility-filtering/ (+2,701/−1,442), phoenix/ (+1,371/−877), home-mixer/ (+1,075/−234), phoenix-rankall/ (+337/−44) and grox/ (+104/−4).Full analysis → Summary: automatic analysis
value_model_gate (537 lines). It was a per-viewer linear gate: it scored features from the viewer's recent action sequence and compared the score to a threshold to decide whether ranking_scorer should use the alternative "dwell regret" scoring path. Inside the hysteresis band, a hash of the user ID made the call. With it go 18 params: every DwellRegret* alpha/negative/gate/temperature param and ValueModelMode. ranking_scorer.rs loses 465 lines and gains 12.PhoenixColdStartMaxResults (default 0, i.e. off). PhoenixRetrievalMOEInferenceClusterId changes from "Experiment2Memy04" to "Experiment3Memy04".UthPctdAccountTakedownEventsJob.scala and underTheHoodLabels.strato.tes_composite hydration path and a batch evaluate_tweets API.Full analysis → Summary: hand-written notes, checked against the diff
task_safety_ptos_media_injected_adult_infrared_video_spam_detection_filter.BRAZIL_2026_ELECTION_USER_IDS (brazil_2026_election_filter.rs) grew from 2,707 to 2,776 entries (per its in-file test assertion).AdInfo.engagement_counts, EngagementCountRow.action, EngagementCountRow.counts, EngagementCounts.present and 8 more).phoenix/ (+423/−81), abuse-enforcement-service/ (+433/−62), grox/ (+489/−6), home-mixer/ (+144/−5) and visibility-filtering/ (+2/−0).Full analysis → Summary: automatic analysis
ColdStartMaxPostAgeSecs: 86400 → 172800.EnableCdwellOnImpr (default off) and EnableFavHoldout (default off).EnableClickDwellLowFavRatePenalty.ClickDwellLowFavRatePenaltyAlpha, ClickDwellLowFavRatePenaltyBaseline, ClickDwellLowFavRatePenaltyCap and ClickDwellLowFavRatePenaltyFloor.fav_holdout_filter and RedactUnenforcedTakedownsFilter.GROK_GEMMA_FOLLOWER_SPLIT 180_000 → 200_000 (constants.py) and FOLLOWER_COUNT_THRESHOLD_FOR_SPAM_DETECTION 30000 → 50000 (task_filter.py).reserved 234 to 252; 214→234 (recsys.proto) and reserved 234 to 252; 214→234 (recsys.proto).visibility-filtering/ (+84/−3,301), under-the-hood/ (+1,477/−137), home-mixer/ (+199/−112), phoenix/ (+148/−72) and takedowns/ (+218/−0).Full analysis → Summary: automatic analysis
EnableAdsBrandSafetyVerdictV2 (enable ads brand safety verdict v2) is now OFF by default.phoenix/ (+437/−38), visibility-filtering/ (+53/−184), grox/ (+38/−7), abuse-enforcement-service/ (+20/−1) and home-mixer/ (+4/−3).Full analysis → Summary: automatic analysis
EnableResponseDiversityStatsExperimentBucket (default off).response_diversity_stats_side_effect.metrics_export_interval_ms 30_000 → 120_000 (kafkaloader.py).FOLLOWER_COUNT_THRESHOLD_FOR_SPAM_DETECTION 10000→30000 (task_filter.py) and export_interval_ms 30_000→120_000 (kafkaloader.py).PredictNextActionsRequest.purchase_value_baselines, PurchaseValueBaseline.advertiser_account_id, PurchaseValueBaseline.impression_id and PurchaseValueBaseline.mean_value_usd_28d).home-mixer/ (+457/−5), phoenix/ (+136/−3) and grox/ (+1/−1).Full analysis → Summary: automatic analysis
BRAZIL_2026_ELECTION_USER_IDS (brazil_2026_election_filter.rs) grew from 2,554 to 2,707 entries (per its in-file test assertion).NextActionDistribution.requestedActionLogits and PredictNextActionsRequest.returnLogitsList).home-mixer/ (+559/−98), phoenix/ (+311/−60), visibility-filtering/ (+282/−68), vm-ranker/ (+153/−41) and grox/ (+60/−55).Full analysis → Summary: automatic analysis
_MOTION_REVEAL_MAX_INPUT_FRAMES 32 → 48 (image.py) and _MOTION_REVEAL_DENSE_SAMPLES 32 → 48 (video_frames.py).("grpc.keepalive_timeout_ms", 20000), 5000→20000 (service.py) and ("grpc.http2.min_ping_interval_without_data_ms", 5 300000→5000 (service.py).UthAccountLabelAggregate.countryCodes and UthPostLabelAggregate.countryCodes).visibility-filtering/ (+1,118/−355), grox/ (+291/−28), under-the-hood/ (+145/−0), phoenix/ (+27/−6) and home-mixer/ (+6/−2).Full analysis → Summary: automatic analysis
home-mixer/ (+452/−107), visibility-filtering-client/ (+137/−62), grox/ (+162/−31), phoenix/ (+42/−8) and candidate-pipeline/ (+3/−40).Full analysis → Summary: automatic analysis
PhoenixRetrievalMOEInferenceClusterId: "Experiment1Fou" → "Experiment2Memy04".WeightPerturbationSalt = "" and WeightPerturbationSigma = 0.0.task_safety_ptos_adult_content_leading_frames_filter.FOLLOWER_COUNT_THRESHOLD_FOR_REPLY_RANKING 120000 → 150000 (task_filter.py), _MOTION_REVEAL_MAX_RESIDUAL_P99 75.0 → 100.0 (image.py), _MOTION_REVEAL_DENSE_SAMPLES 16 → 32 (video_frames.py) and AOT_CACHE_VERSION "20260705" → "20260904" (aot.py).FOLLOWER_COUNT_THRESHOLD_FOR_SPAM_DETECTION 120000→150000 (task_filter.py) and FOLLOWER_COUNT_THRESHOLD_FOR_SPAM_DETECTION 5000→10000 (task_filter.py).ContentFeatures.hasPhoto, ContentFeatures.hasUrl, ContentFeatures.hasVideo, ContentFeatures.maxVideoDurationMs and 5 more).phoenix/ (+8,569/−679), grox/ (+1,535/−74), home-mixer/ (+574/−141), candidate-pipeline/ (+105/−47) and visibility-filtering-client/ (+119/−0).Full analysis → Summary: automatic analysis
PredictNextActionsRequest.experiment_overrides).home-mixer/ (+236/−75), visibility-filtering/ (+128/−61) and phoenix/ (+116/−43).Full analysis → Summary: automatic analysis
PhoenixExperimentOverrides = "".use_pinned_d2h False → True (model_runner.py).FOLLOWER_COUNT_THRESHOLD_FOR_SPAM_DETECTION 1000→5000 (task_filter.py), input_output_aliases 12→14, 13→15, 14→16 (ranker_attention_fa4.py) and default False→True (launch_inference.py).BRAZIL_2026_ELECTION_USER_IDS (brazil_2026_election_filter.rs) grew from 2,388 to 2,554 entries (per its in-file test assertion).visibility-filtering/ (+560/−1,511), phoenix/ (+425/−367), grox/ (+504/−19), home-mixer/ (+408/−29) and abuse-enforcement-service/ (+57/−0).Full analysis → Summary: automatic analysis
Merge of PR #88 from joshs1017dev. VFCandidateHydrator now sorts and dedups in_network_ids (sort_unstable() + dedup()) before calling visibility filtering, so duplicate IDs aren't looked up twice. Two lines changed. This is the only commit in the history from an outside contributor.
Full analysis → Summary: hand-written notes, checked against the diff
PhoenixRetrievalAggregationType: "DENSE_WITH_SHORT_DWELL" → "DENSE_WITH_LONG_DWELL".qk_norm False→True (xrecsys_two_tower.py) and attn_logit_cap 80.0→-1 (xrecsys_two_tower.py).NextActionDistribution.backboneContinuousValues, NextActionDistribution.backboneTopLogProbs and PredictNextActionsRequest.returnBackboneScores).visibility-filtering/ (+392/−426), phoenix/ (+338/−93), scarecrow/ (+333/−0), home-mixer/ (+16/−4) and grox/ (+2/−2).Full analysis → Summary: automatic analysis
AdsBlenderType (feature switch rust_home_mixer_ads_blender) changes from "partition_organic_low_risk" to "multi_risk". That makes the multi-risk blender added on Aug 28 the default for placing ads in the feed. Apart from home-mixer, this sync mostly touches visibility-filtering/rules/, phoenix/ and grox/.
Full analysis → Summary: hand-written notes, checked against the diff
EnablePhoenixScoreStatsExperimentBucket (default off).task_safety_ptos_special_video_filter.BRAZIL_2026_ELECTION_USER_IDS (brazil_2026_election_filter.rs) grew from 2,315 to 2,388 entries (per its in-file test assertion).CandidateDistributionSet.pageDecodeResult, PageDecodeParams.poolN, PageDecodeParams.steps, PageDecodeParams.temperature and 4 more).visibility-filtering/ (+680/−203), phoenix/ (+438/−189), grox/ (+479/−19), home-mixer/ (+313/−94) and abuse-enforcement-service/ (+104/−173).Full analysis → Summary: automatic analysis
RerankerHeadTag = 0.visibility-filtering/ (+1,825/−1,144), phoenix/ (+112/−27), home-mixer/ (+12/−3) and (repo root) (+1/−1).Full analysis → Summary: automatic analysis
MultiRiskAdsBlender (home-mixer/ads/multi_risk_blender.rs). It places ads by the brand-safety verdict of the surrounding organic posts. It counts safe, medium-risk and high-risk posts, caps ad slots at half the safe posts (and half the medium-risk posts for the medium path), keeps the requested ad spacing, and records stats when ads can't be served (no_ads, too_few_posts, …). This commit doesn't make it the default (see Sep 1).viewer_muted_keyword_filter and the response_stats side effect get small additions, and a strato_context helper is added.Full analysis → Summary: hand-written notes, checked against the diff
ConvAssetIds.asset_id, ConvAssetIds.author_id, ConvAssetIds.impressed_time_ms and PredictNextActionsRequest.conv_asset_ids).phoenix/ (+744/−172), visibility-filtering/ (+47/−341), grox/ (+73/−11) and home-mixer/ (+9/−1).Full analysis → Summary: automatic analysis
home-mixer/params/param.rs: DwellWeight 0.0 → 0.05, VideoOpenWeight 0.05 → 0.07, VqvWeight (video quality view) 0.05 → 0.0.PhoenixAggregationType default goes from "DENSE_WITH_SHORT_DWELL" to "DENSE_WITH_LONG_DWELL", and the shadow-traffic cluster list ShadowTrafficPhoenixClusterRates is emptied.EnableAdsBrandSafetyVerdictV2 is now on by default. ColdStartTsTopK drops from 5 to 2.UseServedSlateContext, VMRankerValueModelId, VMRankerSendHeadWeights. EnableMpnScoring is replaced by a new MultiplierPreOffset flag. ranking_scorer.rs is reworked: the per-candidate effective_head_weights function is removed and the author-diversity code is restructured.view_count_on_home (impressions on Home) instead of the overall view_count.following_blocked_by_hydrator flags retweets and quotes whose original author blocks the viewer.Full analysis → Summary: hand-written notes, checked against the diff
FOLLOWER_COUNT_THRESHOLD_FOR_REPLY_RANKING 80000 → 100000 (task_filter.py).FOLLOWER_COUNT_THRESHOLD_FOR_SPAM_DETECTION 80000→100000 (task_filter.py).visibility-filtering/ (+906/−6), grox/ (+294/−56), phoenix/ (+178/−36) and home-mixer/ (+40/−19).Full analysis → Summary: automatic analysis
FOLLOWER_COUNT_THRESHOLD_FOR_REPLY_RANKING 60000 → 80000 (task_filter.py).mode 4.5→4.6 (classifier.py) and FOLLOWER_COUNT_THRESHOLD_FOR_SPAM_DETECTION 60000→80000 (task_filter.py).NextActionDistribution.slateContext).phoenix/ (+835/−79), grox/ (+8/−66) and visibility-filtering/ (+13/−7).Full analysis → Summary: automatic analysis
EnableAiTrendFeedbackContext (default off) and UseServedSlateContext (default off).ai_trend_feedback_context_hydrator._GROK_4_6_INTERNAL_DIAL 0.1 → 0.3 (classifier.py).phoenix/ (+189/−169), home-mixer/ (+255/−17) and grox/ (+35/−14).Full analysis → Summary: automatic analysis
quoted_post_text_hydrator.following_viewer_muted_keyword_filter.muted_keyword_filter → viewer_muted_keyword_filter.MULTIPART_PART_SIZE 16 * 1024 * 1024 → 32 * 1024 * 1024 (checkpoint_store.rs).mode 4.5→4.6 (classifier.py), learning_rate 1e-3→7.1e-4 (xrecsys.py), static_assert(sizeof(UpdateScalars) == 4 * sizeof( 3→4 (async_emb_kernel.hpp), input_output_aliases 12→14, 13→15, 14→16 (ranker_attention_varlen_fa4.py), input_output_aliases 12→14 (ranker_attention_varlen_fa4.py) and 2 more.FetchAggregatedUserActionRequests.feature_switch_overrides, NextActionDistribution.rewardOutputs, RewardOutputs.policyLogits, RewardOutputs.values and 4 more).phoenix/ (+2,811/−536), home-mixer/ (+222/−25), grox/ (+146/−88) and visibility-filtering/ (+124/−2).Full analysis → Summary: automatic analysis
vf_following_candidate_hydrator.FOLLOWER_COUNT_THRESHOLD_FOR_REPLY_RANKING 40000 → 60000 (task_filter.py).FOLLOWER_COUNT_THRESHOLD_FOR_SPAM_DETECTION 40000→60000 (task_filter.py), counts.get(&(1, pb::ServedType::ForYouPhoenixRetri 10→1 (author_cold_start.rs) and counts.get(&(2, pb::ServedType::ForYouPhoenixRetri 10→2 (author_cold_start.rs).home-mixer/ (+185/−42), grox/ (+8/−19), phoenix-rankall-strato/ (+12/−9) and visibility-filtering/ (+3/−0).Full analysis → Summary: automatic analysis
EnableAdsBrandSafetyVerdictV2 (default off).min_actions_for_enforcement 30 → 999999 (score_results_sink_focal.py), FOLLOWER_COUNT_THRESHOLD_FOR_REPLY_RANKING 30000 → 40000 (task_filter.py) and FOLLOWING_MAX_RESULT_SIZE 100 → 110 (config.rs).version 13→14 (sink_policy.yaml), min_actions_for_enforcement 30→999999 (sink_policy.yaml), 2026-08-14: enforcement policy for the 8-head beha 13→14 (sink_policy.yaml) and FOLLOWER_COUNT_THRESHOLD_FOR_SPAM_DETECTION 30000→40000 (task_filter.py).grox/ (+375/−57), home-mixer/ (+276/−15), bdsm/ (+58/−33), phoenix-rankall-strato/ (+35/−44) and phoenix/ (+5/−13).Full analysis → Summary: automatic analysis
Brazil2026ElectionFilter (home-mixer/filters/brazil_2026_election_filter.rs). It removes a candidate when its author, the retweeted author, the quoted author, or any author in the reply's ancestor chain is on a hard-coded list of accounts reported to Brazil's Electoral Court, unless the viewer follows that account. The list starts with 665 user IDs, and code comments cite the TSE open-data candidate dataset. The upstream README's "August 14th, 2026" notes describe the same filter.EnableColdStartThompsonSampling (default off), ColdStartBetaAlpha0 = 0.75, ColdStartBetaBeta0 = 49.25, ColdStartTsTopK = 5, ColdStartImpressionScale = 1.0 and ColdStartTrackedIds. These let the cold-start boost choose among eligible low-impression posts by sampling a Beta posterior of likes per impression, rather than taking the highest-scored one.home-mixer/params/param.rs explain that each weight multiplies the *predicted* probability of an action, not raw engagement counts. The README is restructured into a "Latest Updates" list with an August 14th entry.Full analysis → Summary: hand-written notes, checked against the diff
Deletes the two-line .gitattributes that routed *.npz and *.zip files through Git LFS. No algorithm code changed.
Full analysis → Summary: hand-written notes, checked against the diff
The biggest drop in the repo's history (2,053 files, +363,246/−11,640). The upstream README's "August 13th, 2026" notes cover it.
home-mixer/params/param.rs appears with 189 param! feature-switch defaults, including the action weights that blend predicted engagement probabilities into one score (for example ReplyWeight, ShareViaCopyLinkWeight and the negative ReportWeight/BlockAuthorWeight). See the parameter tracker on this site for how they've changed since.visibility-filtering/ decides whether a post is shown, dropped, or shown behind an interstitial.botmaker/, botmaker-rules/, scarecrow/), account-scoring models (agatha/, bdsm/, user-cred-v2/), media models (media-model-proxy/, clip/) and enforcement (abuse-enforcement-service/).simclusters/) is added as another out-of-network candidate source, alongside Thunder and Phoenix retrieval.ad_adjacent_served_filter, following_retweet_deduplication_filter, self_reply_chain_filter and oon_nsfw_simclusters_filter. See the full analysis for the complete list.docs/BIDIRECTIONAL_BOOST_CHANGE.md is a worked example of a changelog. It walks through the July 2026 rollout of the bidirectional-follow reply boost (A/B tested at 5/10/15/20 from July 10, broadly launched at 20 on July 13, set to 15 on July 24) and says future updates will be published to the repo so readers "should be able to understand what changed by checking the diffs."Full analysis → Summary: hand-written notes, checked against the diff
The only change is the Git LFS pointer for phoenix/artifacts/oss-phoenix-artifacts.zip. The archive's recorded size drops from 3,123,149,995 to 2,903,518,802 bytes. No source code changed.
Full analysis → Summary: hand-written notes, checked against the diff
A large update (187 files, +18,263/−926). The upstream README gained an "Updates — May 15th, 2026" section for this release, which is no longer in the current README.
home-mixer/sources/: ads_source, cached_posts_source, phoenix_moe_source, phoenix_topics_source, prompts_source, push_to_home_source, scored_posts_source, tweet_mixer_source and who_to_follow_source.ranking_scorer and vm_ranker (the original weighted_scorer, oon_scorer and author_diversity_scorer files are still present).ancillary_vf_filter, new_user_topic_ids_filter, previously_seen_posts_backup_filter, topic_ids_filter and video_filter.blender_selector, plus a new home-mixer/ads/ module for placing ads in the feed.grox/ content-understanding service (classifiers, embedders, task engine) and many new query/candidate hydrators.run_pipeline.py entry point and a pre-trained mini model distributed through Git LFS (phoenix/artifacts/oss-phoenix-artifacts.zip, about 3.1 GB).Full analysis → Summary: hand-written notes, checked against the diff
The first commit in the repository's history: 79 files across candidate-pipeline/, home-mixer/, phoenix/ and thunder/.
home-mixer/ is the service that assembles the For You feed. It ships with 12 filters (e.g. age_filter, author_socialgraph_filter, muted_keyword_filter, previously_seen_posts_filter, vf_filter), two candidate sources (thunder_source for in-network posts, phoenix_source for out-of-network retrieval) and four scorers: phoenix_scorer, weighted_scorer, author_diversity_scorer and oon_scorer.thunder/ holds recent posts from followed accounts in memory; phoenix/ contains the retrieval/ranking model code.candidate-pipeline/ is the generic source → hydrate → filter → score → select framework the feed is built on.Full analysis → Summary: hand-written notes, checked against the diff