Lesson 3

How Posts Actually Get Scored

Phoenix predicts probabilities for many actions — not one relevance score

The Scoring Model

Phoenix doesn't produce a single "relevance score." It predicts the probability that you'll take each of several specific actions on a post:

ActionWeight (as of 2026-08-14)What it means
Favorite0.5You'll like the post
Reply5.0You'll reply to the post
Retweet1.0You'll repost it
Quote5.0You'll quote it
Share (via DM)5.0You'll DM it to someone
Share (via copy link)20.0You'll copy the link
Profile click0.0You'll click the author's profile
Follow author4.0You'll follow the author
Dwell time0.0How long you spend on the post
Not interested-43.2You'll mark "not interested"
Block author-31.2You'll block the author
Mute author-58.8You'll mute the author
Report-234.0You'll report the post

The Final Score

A weighted sum of these predicted probabilities produces the final ranking score:

score = Σ (weight_i × P(action_i))

Critical Correction (Aug 14, 2026)

The repo explicitly warns: these weights multiply predicted probabilities, not raw engagement counts. It's incorrect to say "1 report cancels out 468 likes" — the weights scale *your predicted likelihood* of each action, not the count of actions taken.

The baseline probability of a Report is more than 1000x lower than a Like, so it's weighted heavily to ensure the prediction can actually affect ranking. But a report from a bad actor primarily affects recommendations for users *similar to that bad actor*, not the post's ranking for everyone.

Check Your Understanding

1. What does Phoenix predict for each post?
  1. A. A single relevance score
  2. B. Probabilities for many separate actions✓ correct
  3. C. A sentiment score
  4. D. A quality rating

Why: Phoenix predicts the probability of each action (favorite, reply, repost, quote, dwell, block, report, etc.) — not one score.

2. What is the base weight of a Reply vs a Favorite?
  1. A. Reply is 2x a Favorite
  2. B. Reply is 10x a Favorite (5.0 vs 0.5)✓ correct
  3. C. They are equal
  4. D. Reply is 100x a Favorite

Why: ReplyWeight = 5.0, FavoriteWeight = 0.5 — a reply is weighted 10x a favorite in the base config.

3. Is it correct to say "1 report cancels out 468 likes"?
  1. A. Yes, that's exactly right
  2. B. No — weights multiply predicted probabilities, not raw counts✓ correct
  3. C. Only for verified accounts
  4. D. It depends on the time of day

Why: The repo explicitly warns against this interpretation. Weights scale predicted probabilities, not raw engagement counts.

Exercise

Look at your last 5 posts. Which ones got replies vs only likes? The ones with replies were weighted ~10x higher in your followers' feeds. This is why reply-bait works.

Hint: A post that gets 10 likes and 0 replies scores lower than a post that gets 2 likes and 1 reply.

How Posts Actually Get Scored | Qubax Growth Lab · Qubax AI