Skip to content

Stop Phoenix authorIsNsfwSeq from skipping retweets - #73

Closed
Pitchfork-and-Torch wants to merge 19 commits into
mainfrom
cursor/magellan-overnight-b-nsfw-9cba
Closed

Stop Phoenix authorIsNsfwSeq from skipping retweets#73
Pitchfork-and-Torch wants to merge 19 commits into
mainfrom
cursor/magellan-overnight-b-nsfw-9cba

Conversation

@Pitchfork-and-Torch

Copy link
Copy Markdown
Owner

Phoenix hashes get_original_author_id() but as_tweet_info zeroed safety_label_mask on every retweet. Gizmoduck already fetched the origin user and then wrote the retweeter's NSFW bit into nsfw_author_phoenix. Training reads safetyLabelMaskSeq bit 2; serving therefore scored every RT of an NSFW original as a clean-author post.

Proof

  • Entry: GizmoduckCandidateHydrator (origin user already fetched) then PostCandidate::as_tweet_info
  • Sink: PhoenixScorer -> InputBuffer authorIsNsfwSeq from TweetInfo.safety_label_mask bit 2
  • Break: retweeted_user_id.is_none() guard forced mask 0; phoenix bit keyed the retweeter
  • Viewer: in-network RTs of NSFW originals rank as if the origin author were clean
  • Twin: tweet_id and author_id already use the origin; originals already stamp nsfw_author_phoenix; history UAS stamps the engaged tweet mask

This is not xai-org#125 (size residual, quoted OR, gizmoduck fail-closed). Not xai-org#185 (OON NSFW pre-score filter). Not xai-org#172/xai-org#188 (diversity/MoE retweeter key). Not xai-org#173 (safemodel tweet labels). Not xai-org#189 (DedupConversationFilter). Filter and ads poster bits are unchanged.

Fix

  • nsfw_author_phoenix on a retweet is the origin author's bit
  • as_tweet_info stamps that bit on retweets too

Tests cover origin-NSFW RT stamp, retweeter-NSFW clean-origin no stamp, and original-post unchanged.

cargo test cannot run home-mixer. Public dump has no Home Mixer manifest. Standalone rustc model of old vs new stamp: 5/5 passed.

Fork PR: none

Open in Web Open in Cursor 

CI agent and others added 19 commits August 14, 2026 20:55
in_network_ids is passed to the VF client without deduplication, while
oon_ids is deduped four lines below. retweeted_tweet_id is pushed for
every candidate that has one, so the same ID repeats once per retweet of
a given post — most often when that post is going viral.

Neither VfClient implementation dedupes its input: StratoVfClient builds
one call per element, and XaiVfClient chunks by XAI_VF_MAX_BATCH_SIZE, so
duplicates consume batch slots and can force an extra round trip.

Not a correctness issue — results collapse into a HashMap keyed by tweet
ID — but redundant work on the For You serving path.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Deduplicate in_network_ids before VF lookup
as_tweet_info zeroed safety_label_mask on every retweet, so Phoenix
hashed the origin author with authorIsNsfwSeq=0. Gizmoduck already
fetched the origin user but wrote the retweeter's NSFW bit.

Stamp nsfw_author_phoenix from the origin on retweets and drop the
retweet skip in as_tweet_info. Filter/ads poster bits are unchanged.

Co-authored-by: Jon Bailey <Pitchfork-and-Torch@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants