Reply — a positive ranking signal
Predicted probability the viewer replies. 27× a like in the 2023 stack; the 2026 weight is withheld but the head exists and is separately modeled.
Last public weight (2023 heavy ranker): 13.5 LEGACY-2023 — from the frozen twitter/the-algorithm; absent from the 2026 code.
2026 weight: withheld — all current weights live in an
unpublished params module. The weights story →
What actually moves it
Posts that leave something to answer — a real question, a defensible claim, a gap the reader can fill — earn this head. Note Grox separately ranks replies and screens reply-spam, so low-effort bait cuts the other way.
Where it sits in the formula
Phoenix (a Grok-1-architecture transformer, phoenix/README.md)
predicts P(reply) per viewer from that viewer’s own
engagement history — not from your post’s running like-count. The prediction is
multiplied by its weight and summed with the other 21 terms, then the
author-diversity decay and
out-of-network multiplier apply.
Related: Favorite (Like) · Repost · Photo expand · Click (open post)