2026-10-30

1Serving page log turned on and ads V2 verdict turned off — X

• The serving page log switch changed from false to true by default verified — params/param.rs:312-317.
  • The switch is named EnablePageLogging (rust_home_mixer_enable_page_logging), and the default TTL of 3600 seconds is unchanged.
  • This log is used only in the production environment. The condition is is_prod() && EnablePageLogging — phoenix_request_cache_side_effect.rs:304-306.
• The log contains the first posts of the feed actually served, in the order served, with their weighted-sum scores and final scores verified — same file:201-277.
  • The two slot probability fields (reward_rerank_slot_prob, page_decode_slot_prob) are filled only when the exploration or decoder switch is on.
  • The two switches (EnablePageRlExploration, EnablePageDecodeRerank) are still false by default in this commit. So the logs accumulating now hold only the order and scores, without probabilities.
• X appears to have started by recording the pages served by the existing policy before turning on RL re-ranking inferred. The evidence is that, of the three switches, only the log was turned on first.

• The switch that judges the risk level of posts next to ads under V2 rules changed from true to false verified — params/param.rs:938-943.
  • The switch is named EnableAdsBrandSafetyVerdictV2 (rust_home_mixer_ads_bs_v2_exp_enabled).
  • Its history: introduced off on 2026-08-17, on 08-25, off 09-16, on 10-03, off 10-29.
• The fallback switch that fills gaps with an author-level safety label did not change verified — the same diff has no line for it.
• The reason for turning it off is not in the code inferred. The switch key contains exp, so it appears to be a value that opens and closes an experiment bucket, and four changes in two months fit that reading.

• slot_spread and least_movement were added to the ads blender choices, and the default is still multi-risk verified — selectors/blender_selector.rs:62-67.

  • The param.rs diff of this commit has no AdsBlenderType line.

• Both blenders find two organic posts that can sit above and below each ad verified — ads/neighbours.rs.

  • Posts fall into four grades: Safe, Low, Medium, and High. A High post cannot sit next to any ad, and a Medium post can sit only next to an ad whose advertiser allows that context (BSR-high).
  • A post containing a handle or an excluded keyword specified by the advertiser also cannot be a neighbor.

• slot_spread first sets ad slots evenly across the whole page, then pulls the nearest safe posts above and below each slot verified — ads/slot_spread_blender.rs:1293-1303.

  • It keeps the minimum gap and divides the remaining positions equally between slots.
  • If sensitive ads remain for later, it uses Low posts first to save Safe posts.

• least_movement picks the placement that disturbs the original order of organic posts the least verified — ads/least_movement_blender.rs.

  • When deciding whether to accept an ad, it checks by bipartite matching (augmenting path) whether every ad, including those already chosen, can be given two non-overlapping neighbors.
  • When comparing slot candidates, it uses a slack weight of 50 (FIRST_AD_SLACK_WEIGHT) for the first ad and 2 (SLACK_WEIGHT) for later ads. This appears to be a value meant to strongly avoid placements that delay the first ad inferred.

• The tests pin that both blenders keep the same safety conditions as multi-risk and do not serve fewer ads than multi-risk verified — ads/tests/blender_comparison_test.rs.

  • On the same mixed feed, both serve 7 ads and drop 3 for the gap limit, so the results are the same. The number of organic posts moved is 6 for least_movement and 8 for slot_spread.

  assert!(
      total >= baseline_total,
      "{name} served {total} ads vs multi_risk {baseline_total}"
  );
  

• The ads lead-generation conversion (ADS_LEADGEN_CONVERSION) entered a new head group that trains "only on clicks, in the fresh window" verified — ads_head_masking.py.
  • This action was added to the proto in the previous day's commit. The existing web and app conversion heads train in the delayed window (the window that waits for late-arriving conversions).
  • The training mask is fresh_ct & fresh & kept & click.
• The page RL training config gained a variant that switches the dwell loss to CREAD (a method that trains binary classifiers for whether dwell time exceeded several thresholds, then recovers the expectation) verified — configs/xrecsys.py.
  • It is named immersive_page_rl_cread. It sets the Tweedie and MAE weights to 0, sets the two CREAD losses to 0.1 each, and turns on the dwell bucket table.
  • The default home config already switched to CREAD on 2026-10-08. Now the page RL side can use the same loss.

• The video carousel ranking score gained a "probability of opening the video × expected dwell time on the immersive surface" term, with a weight defaulting to 0.0 verified — util/video_carousel.rs:178-190.
• Carousel module items were removed from the recent impression history, so carousel videos are not counted as "already seen posts" of the main feed verified — served_history_query_hydrator.rs:94-110. The carousel is off by default, so default behavior is unchanged.
• An option that also updates the user token in the candidate-only layer (the last Transformer layer that updates only candidate tokens) was introduced, off by default, so that page RL can run with that layer verified — models/transformer.py.
• An option that restricts ads head metrics to the window each head trains on was introduced, off by default verified — models/recsys_model.py.
• Metrics that count ad placement results and supply limits were attached to the two new blenders verified — ads/health_metrics.rs.
• VF (visibility-filtering, a service that decides for each post whether this viewer may see it) is a refactor. The rules, the per-safety-level rule lists, and the client switch defaults did not change verified — 606b254.
  • It merged quote-relationship lookup into the Flock source, fixing a problem where answers got mixed when the viewer was the quote author.
  • It returns posts it cannot find as NotFound instead of Failed. Either way, the post is not served.
  • The golden corpus (tests of expected verdicts per rule) gained cases for the For You recommendation and immersive recommendation surfaces.
• The ranker cluster label was removed from training statistics, and the multimodal embedding lookup uses only the O2 store with no fallback path verified — 606b254.
• grox (the Grok-based annotation pipeline) gained an affiliate-company badge image field and a helper that extracts JSON from LLM output verified — same commit.

What to check next