2026-11-12

1Two safety and dedup rules for the video carousel are now on by default — X

• The default of the switch that removes posts with a brand safety MediumRisk verdict from carousel candidates changed from false to true verified — params/param.rs:1354-1359.
  • The switch is named rust_home_mixer_enable_video_carousel_medium_risk_filter and was introduced as off by default in the 2026-11-03 commit (Gen4+55).
  • The filter code removes a post with no verdict if either the HighRisk exclusion or the MediumRisk exclusion is on. So the handling of posts with no verdict is the same as before.
• The filter that does not re-insert videos seen through the carousel in the last 3 days is also on by default verified — params/param.rs:1360-1365.
  • The switch is named rust_home_mixer_enable_video_carousel_seen_filter and was introduced as off by default in the 2026-11-06 commit (Gen4+58). The seen record is read from a per-user Redis.
• The switch that turns on the carousel itself is still off by default in this commit verified — params/param.rs:1282-1287.

• With the switch on, the carousel candidate pool starts from main feed posts below the carousel slot, and videos picked for the carousel are removed from the main feed verified — filters/video_carousel_filter.rs.

  • Until now the candidate pool was posts beyond the main feed result size (RESULT_SIZE) plus carousel-only extra candidates. When the switch is on, the starting point becomes VideoCarouselPosition − 1 (default 5).
  • When removing from the main feed, a post counts as the same video if any one of the tweet ID, the video media ID or the MD5 hash matches the carousel video — util/video_carousel.rs.
  • The switch is named rust_home_mixer_enable_video_carousel_take_from_feed and defaults to false — params/param.rs:1366-1371.

• The test pins that the 11th video in the main feed moves to the carousel and that the 21st post, which uses the same media, is also removed from the main feed verified — same file.

  assert_eq!(ids(query.video_carousel.get().unwrap()), [11, 101, 102, 103, 104]);
  assert_eq!(kept[..5], [1, 2, 3, 4, 5]);
  assert!(!kept.contains(&11) && !kept.contains(&21));
  

• A Grok suggested-tasks card was added to the frame list (frames are non-post cards inserted between feed items, such as sports cards), and its slot is third in the Home feed verified — frames/catalog.rs.
  • The path is /grokbot/suggested_tasks; until now frames existed only on soccer, NFL and MLB topic timelines.
  • The card appears only in the first response, not in new-post check requests (polling) or in requests that scroll down for more. The user cannot dismiss it.
  • It does not appear in Topic requests or in the Following feed.
• After the card has been shown once, it reappears only after the fatigue time (default 6 hours) has passed verified — params/param.rs:1256-1267.
  • The switches are rust_home_mixer_jetfuel_frame_grokbot_suggested_tasks_enabled (default false) and ..._fatigue_hours (default 6).
  • The card content is received directly by viewer ID from a Jetfuel client (Jetfuel is an internal service that returns the rendered content of frame cards), not from Strato — sources/jetfuel_direct_frame_source.rs.

• When on, up to 64 posts the viewer recently viewed are chosen in order of impression time and attached, together with topics, to the VMRanker (slate reordering service) request verified — query_hydrators/recent_views_query_hydrator.rs.
  • For each viewed post it sends the impression time, surface, action type bits and dwell time (ms). The history is received as an aggregation type that includes recent impressions (DENSE_WITH_RECENT_IMPRESSIONS), up to 256 rows.
  • It does not run for Following requests or Topic requests.
• On the candidate side, the topics of up to 200 top candidates by local weighted-sum score are looked up within 80ms and attached to each candidate verified — scorers/vm_ranker.rs.
  • The switch is rust_home_mixer_vm_ranker_send_post_topics (default false), and the caps are PostTopicsMaxCandidates = 200, PostTopicsTimeoutMs = 80 and RecentViewsRows = 64 — params/param.rs:385-414.
  • If the lookup times out, the request is sent without topics.
• How VMRanker uses these values is again not in the public code verified. We infer this is preparation for testing a diversity objective that compares the "just viewed" topics with candidate topics inferred. The basis is that it uses the same experiment payload path as the history topic sending in the 10-20 commit.

• The safety label cache, which after filling re-read the repository twice at 30 and 90 seconds, now reads it once at 60 seconds, and a metric counts whether the re-read result changed verified — safety-label-source/dragonfly.rs.
  • The default retention time for a no-label result (not_found_ttl) increased from 10 minutes to 1 hour.
• A post label action was added to the action settings of abuse-enforcement-service (a service that takes account and post actions based on model scores), and the selectable labels are NsfwHighPrecision and SoftNsfw verified — types/src/action_control.rs.
• Experiment20Fou was removed from the shadow traffic targets (a path that sends only scoring to an experimental Phoenix cluster and is not used for serving), and Experiment24~28Fou were added at a 0.3 ratio each verified — params/param.rs:91-106.
• The TwHIN post and video embedding streaming job first filters for only like and video quality view (VQV) events, then writes to two paths. The action types each path receives are unchanged, so the result is the same verified — TwhinTweetJob.scala.
• Several filters were changed to use the same split helper, and topic annotation decoding was moved to a shared util. There is no policy change verified — commit diff.
• Two unused history aggregation types (33, 35) were changed to reserved numbers in the proto, and a few help-link constants in the VF client were deleted verified — same diff.
• The last sync time of the parameter mirror is 2026-11-10T16:00:31Z verified — params/param.rs:1.

What to check next