The X Scoring System Explained

How your posts are ranked in the For You feed, based on the actual algorithm code.

---

The Scoring Formula

Every post is scored using this formula:

Final Score = Σ (weight_i × P(action_i))

Where:

• P(action_i) = Probability of user taking action (predicted by ML model)
• weight_i = Importance weight for that action

Example:

Score = (0.3 × P(like)) + (0.6 × P(reply)) + (0.4 × P(retweet)) + ...
                         - (0.5 × P(block)) - (1.0 × P(report))

---

The 19 Predicted Actions

The Phoenix ML model predicts probability for each action:

Positive Actions (Increase Score)

Action	Description	Relative Weight
favorite	Like the post	⭐⭐
reply	Reply to the post	⭐⭐⭐ (Highest)
retweet	Retweet	⭐⭐
quote	Quote tweet	⭐⭐⭐
share	General share	⭐
share_via_dm	Share via DM	⭐
share_via_copy_link	Copy link	⭐
click	Click on post	⭐
profile_click	Click author’s profile	⭐⭐
follow_author	Follow after seeing post	⭐⭐⭐
photo_expand	Expand photo	⭐
video_quality_view	Watch video (quality)	⭐⭐
dwell	Stop and read	⭐
dwell_time	Time spent reading	⭐ (continuous)
quoted_click	Click on quoted content	⭐

Negative Actions (Decrease Score)

Action	Description	Relative Weight
not_interested	“Not interested” button	❌
mute_author	Mute the author	❌❌
block_author	Block the author	❌❌❌
report	Report the post	❌❌❌❌ (Devastating)

---

The ML Model: Phoenix

How It Works

┌─────────────────────────────────────────────────────────────────┐
│                         PHOENIX SCORER                           │
├─────────────────────────────────────────────────────────────────┤
│                                                                  │
│  INPUT:                                                          │
│  ├── User embedding (who is viewing)                             │
│  ├── User history (128 recent engagements)                       │
│  └── Candidate posts (up to 32)                                  │
│                                                                  │
│  PROCESSING:                                                     │
│  └── Grok-based Transformer                                      │
│      ├── Attention mechanism                                     │
│      └── Candidate isolation (posts can't see each other)        │
│                                                                  │
│  OUTPUT:                                                         │
│  └── P(action) for each of 19 actions                            │
│                                                                  │
└─────────────────────────────────────────────────────────────────┘

Candidate Isolation

Key insight: Your post’s score doesn’t depend on what other posts are in the batch.

This means:

• Scores are consistent and cacheable
• You can’t “hide” behind other posts
• Each post is judged independently

---

The Scoring Pipeline

Step 1: Weighted Scorer

// From weighted_scorer.rs
combined_score = 
    apply(favorite_score, FAVORITE_WEIGHT)
  + apply(reply_score, REPLY_WEIGHT)        // ~2× other weights
  + apply(retweet_score, RETWEET_WEIGHT)
  + apply(quote_score, QUOTE_WEIGHT)
  + apply(share_score, SHARE_WEIGHT)
  + apply(click_score, CLICK_WEIGHT)
  + apply(profile_click_score, PROFILE_CLICK_WEIGHT)
  + apply(vqv_score, vqv_weight)            // Only if video > min duration
  + apply(dwell_score, DWELL_WEIGHT)
  + apply(dwell_time, CONT_DWELL_TIME_WEIGHT)
  + apply(follow_author_score, FOLLOW_AUTHOR_WEIGHT)
  // Negative weights
  + apply(not_interested_score, NOT_INTERESTED_WEIGHT)  // Negative
  + apply(block_author_score, BLOCK_AUTHOR_WEIGHT)      // ~-10×
  + apply(mute_author_score, MUTE_AUTHOR_WEIGHT)        // ~-5×
  + apply(report_score, REPORT_WEIGHT)                  // ~-20×

Step 2: Author Diversity Scorer

Penalizes multiple posts from same author in one feed:

// From author_diversity_scorer.rs
multiplier = (1 - floor) × decay^position + floor

// Example with decay=0.8, floor=0.2:
// Post 1: 1.0 (100%)
// Post 2: 0.84 (84%)
// Post 3: 0.67 (67%)
// Post 4: 0.54 (54%)
// ...converges to 0.2 (20%)

Step 3: OON Scorer

Penalizes out-of-network content:

// From oon_scorer.rs
if !in_network {
    score *= OON_WEIGHT_FACTOR  // Less than 1.0
}

---

Practical Implications

What This Means for You

Algorithm Behavior	Your Strategy
Replies weighted highest	Create discussion-starting content
Blocks are -10× likes	Avoid controversial content that triggers blocks
Author diversity penalty	Space posts, use threads
OON penalty	Build quality followers first
Video needs min duration	Make videos 10+ seconds
Dwell time tracked	Create longer, engaging content

The Score Optimization Hierarchy

1. Maximize P(reply)     → Ask questions, create debate
2. Maximize P(quote)     → Create quotable, shareable insights
3. Maximize P(follow)    → Provide unique value
4. Maximize P(retweet)   → Make content worth sharing
5. Maximize P(like)      → Be likeable (baseline)
6. Minimize P(block)     → Don't annoy people
7. Minimize P(report)    → Stay within guidelines

---

Source Files

Component	File
Weighted scoring	home-mixer/scorers/weighted_scorer.rs
Author diversity	home-mixer/scorers/author_diversity_scorer.rs
OON penalty	home-mixer/scorers/oon_scorer.rs
Phoenix ML	home-mixer/scorers/phoenix_scorer.rs
Grok transformer	phoenix/grok.py
Ranking model	phoenix/recsys_model.py

---

Next: Content Optimization →