I had Claude clone this repo and dig through the actual code to fact-check this post's claims.
Result:
Mostly crap. There is a continuous per-position decay within a feed response, but no "4/day" tripwire.
In other words, the more you post per day, the more "saturated" it gets. But nothing in the code says four.
More details:
[ claims that are FALSE ]
- "media gets 2x signal weight"...
no multiplier exists. has_media is just a model feature
- "replies weighted by WHO replies" ...
single ReplyWeight in ranking_scorer.rs. who-replied is a UI facepile, not a scorer
- "long-form 4000-char heavier weight"...
zero hits for length/article/char_count anywhere
- "consistent cadence > volume" ...
zero hits for cadence/regularity/posting_frequency
- "self-reply in first 30 min = ranking gold"...
no self-reply logic. SelfTweetFilter literally removes your own posts
[ claims that are UNSUPPORTED (invented) ]
- recycled-template classifier ... doesn't exist
- originality / AI-roundup classifier ... doesn't exist
- engagement-bait closer detector ... doesn't exist
- motivational-fluff penalty ... doesn't exist
- text+media interaction term ... doesn't exist
real classifiers in grox/classifiers/content/: spam, safety_ptos, reply_ranking, banger_initial_screen.
That's it.
[ claims that are PARTIALLY true but distorted ]
- "4+ posts/day dilution" → real mechanism is AuthorDiversityScorer: continuous per-position decay within ONE feed response. no "4/day" tripwire
- "OON discovery 3x'd" → OonWeightFactor exists in ranking_scorer.rs:225. the actual number lives in an external feature-switch service NOT in the repo.
"3x" is made up
- "engagement pods reweighted" → MutualFollowJaccardHydrator exists but emits stats, isn't wired into any scorer
- "model reads full thread context" → conversation_id is tracked, ranker attends over user history + candidate. no explicit full-thread feature
[ score ]
0/14 cleanly true. 4 partially true. 10 false or invented.
Thank you for coming to my Ted Talk
