The Phoenix weighted scorer assigns weight ×1.5 to replies and ×1.2 to quote posts — far above likes (×0.5). Content that sparks conversation is what the algorithm values most.

P(block) has weight −3.0, P(report) has weight −4.0. A handful of blocks can negate dozens of likes. Provocative content that triggers blocking will suppress your overall author score over time.

Thunder trims posts past the retention window entirely — they never enter the candidate pool. Posts older than ~48h rarely compete. Timing your posts when your audience is active is non-negotiable.

P(video_view) and P(dwell) are independent prediction targets. Videos that get watched fully generate compounding positive signals. Autoplay skips actively hurt your score.

The author diversity scorer attenuates repeated-author candidates. If you post 20 times a day, your posts compete with each other for the single author slot per viewer feed. Quality beats quantity.

To appear in non-followers' feeds, your post's embedding (Phoenix two-tower model) must match what those users have engaged with before. Niche-consistent posting over time builds a stronger, more discoverable embedding than posting about everything.

The PreviouslySeenPostsFilter and PreviouslyServedPostsFilter remove posts a viewer has already encountered. Viral spread depends entirely on reaching new viewers — the algorithm doesn't recirculate.

xAI explicitly removed all hand-engineered features. Verified status, follower count, and paid tiers provide no direct scoring boost. Only your historical engagement pattern, as understood by the Grok transformer, drives your score.