diff --git a/TWITTER_ALGORITHM_MASTERY_GUIDE.md b/TWITTER_ALGORITHM_MASTERY_GUIDE.md
new file mode 100644
index 000000000..8bb0a359a
--- /dev/null
+++ b/TWITTER_ALGORITHM_MASTERY_GUIDE.md
@@ -0,0 +1,542 @@
+# Twitter/X Algorithm Mastery Guide
+## A Deep Technical Analysis for Account Growth
+
+*Based on comprehensive analysis of Twitter's open-sourced recommendation algorithm*
+
+---
+
+## Table of Contents
+1. [Executive Summary](#executive-summary)
+2. [How Your Tweets Get Recommended](#how-your-tweets-get-recommended)
+3. [The Scoring Formula Decoded](#the-scoring-formula-decoded)
+4. [Engagement Signals: What Matters Most](#engagement-signals-what-matters-most)
+5. [Positive Signals (Ranked by Impact)](#positive-signals-ranked-by-impact)
+6. [Negative Signals (Avoid These)](#negative-signals-avoid-these)
+7. [The Four Pillars of Recommendation](#the-four-pillars-of-recommendation)
+8. [Actionable Optimization Strategies](#actionable-optimization-strategies)
+9. [API-Based Analysis Opportunities](#api-based-analysis-opportunities)
+10. [Technical Deep Dives](#technical-deep-dives)
+
+---
+
+## Executive Summary
+
+Twitter's recommendation algorithm operates through a **multi-stage pipeline** that narrows ~1 billion tweets down to ~1,500 candidates, then ranks them using ML models that predict **17 different engagement types**. Your final score is a weighted sum of these predictions.
+
+**Key Insight**: The algorithm favors content that generates **"quality" engagement** - not just likes, but replies that get responses from authors, profile visits that lead to follows, and video views with high completion rates.
+
+---
+
+## How Your Tweets Get Recommended
+
+### The Pipeline (4 Stages)
+
+```
+Stage 1: Candidate Retrieval (~1B → ~1,500 tweets)
+├── In-Network (50%): Tweets from accounts you follow
+│   └── Fetched via Earlybird search index (600 tweets)
+├── Out-of-Network (50%): Tweets from accounts you don't follow
+│   ├── UTEG - "Liked by" (300 tweets): What your network liked
+│   ├── SimClusters (200 tweets): Community/topic matching
+│   ├── TwHIN (200 tweets): Knowledge graph embeddings
+│   └── FRS (100 tweets): Follow recommendations
+└── Backfill: Reverse chronological fallback
+
+Stage 2: Feature Hydration (~6,000 features per tweet)
+├── Author features (reputation, follower/following ratio)
+├── Tweet content features (text, media, hashtags)
+├── Engagement features (real-time counts)
+├── Relationship features (RealGraph scores)
+└── Embedding features (SimClusters, TwHIN)
+
+Stage 3: ML Scoring (Navi model)
+├── Predicts probability of 17 engagement types
+├── Each prediction multiplied by configurable weight
+└── Weighted sum = raw score
+
+Stage 4: Heuristic Rescoring
+├── Out-of-network penalty: 0.75x multiplier
+├── Reply penalty: 0.75x multiplier
+├── Author diversity decay: 0.5x for repeated authors
+├── Feedback fatigue: Penalizes recently shown content
+└── Content diversity rescoring
+```
+
+---
+
+## The Scoring Formula Decoded
+
+### Base Score Calculation
+
+```
+Final Score = Σ(predicted_engagement[i] × weight[i]) + ε
+
+Where ε = 0.001 (prevents zero scores)
+```
+
+### The 17 Engagement Predictions
+
+| Engagement Type | What It Means | Relative Impact |
+|-----------------|---------------|-----------------|
+| **Favorite (Like)** | User likes the tweet | HIGH |
+| **Retweet** | User retweets | HIGH |
+| **Reply** | User replies | MEDIUM |
+| **Reply Engaged by Author** | User replies AND author engages back | VERY HIGH |
+| **Good Click V1** | Click → dwell → fav/reply | VERY HIGH |
+| **Good Click V2** | Click → 2+ actions taken | VERY HIGH |
+| **Good Profile Click** | Click profile → engage with profile | HIGH |
+| **Video Playback 50%** | Watch ≥50% of video | HIGH |
+| **Tweet Detail Dwell 15s** | View tweet detail for 15+ seconds | HIGH |
+| **Profile Dwell 20s** | View profile for 20+ seconds | MEDIUM |
+| **Bookmark** | User bookmarks tweet | MEDIUM |
+| **Share** | User shares tweet | HIGH |
+| **Share Menu Click** | User clicks share menu | LOW |
+| **Negative Feedback V2** | User clicks "Not interested" | VERY NEGATIVE |
+| **Report** | User reports tweet | EXTREMELY NEGATIVE |
+| **Strong Negative Feedback** | Block, mute, or report | EXTREMELY NEGATIVE |
+| **Weak Negative Feedback** | See fewer, don't like | NEGATIVE |
+
+### Weight Ranges (from code)
+
+```scala
+// Positive engagement weights: 0 to +10,000
+ModelWeights.FavParam (min: -10000, max: 10000)
+ModelWeights.RetweetParam (min: -10000, max: 10000)
+ModelWeights.ShareParam (min: -10000, max: 10000)
+
+// Negative engagement weights: -20,000 to 0
+ModelWeights.ReportParam (min: -20000, max: 0)
+ModelWeights.StrongNegativeFeedbackParam (min: -1000, max: 0)
+ModelWeights.WeakNegativeFeedbackParam (min: -1000, max: 0)
+```
+
+---
+
+## Engagement Signals: What Matters Most
+
+### Tier 1: Maximum Impact Signals
+
+1. **Reply Engaged by Author** (`PredictedReplyEngagedByAuthorScoreFeature`)
+   - When someone replies AND YOU REPLY BACK
+   - This is a multiplier on engagement quality
+   - **Action**: Always respond to replies, especially within first hour
+
+2. **Good Click** (`PredictedGoodClickConvoDescFavoritedOrRepliedScoreFeature`)
+   - User clicks your tweet → views detail → favorites OR replies
+   - Indicates content that hooks and converts
+   - **Action**: Write compelling first lines; deliver value in thread
+
+3. **Video Playback 50%** (`PredictedVideoPlayback50ScoreFeature`)
+   - User watches ≥50% of your video
+   - Stronger signal than just a like
+   - **Action**: Front-load value; keep videos concise; use captions
+
+### Tier 2: High Impact Signals
+
+4. **Tweet Detail Dwell 15s** (`PredictedTweetDetailDwellScoreFeature`)
+   - User spends 15+ seconds viewing your tweet detail
+   - Indicates substantive, valuable content
+   - **Action**: Write longer-form content worth reading
+
+5. **Share** (`PredictedShareScoreFeature`)
+   - User shares your tweet (DM, copy link, external)
+   - Strong viral signal
+   - **Action**: Create shareable insights, memes, or data
+
+6. **Good Profile Click** (`PredictedGoodProfileClickScoreFeature`)
+   - User clicks profile → engages with profile (follow, like, etc.)
+   - Indicates you're building reputation
+   - **Action**: Maintain consistent, quality profile presence
+
+### Tier 3: Standard Signals
+
+7. **Favorite** - Basic engagement
+8. **Retweet** - Distribution signal
+9. **Reply** - Conversation starter
+10. **Bookmark** - Value indicator (save for later)
+11. **Profile Dwell 20s** - Extended profile interest
+
+---
+
+## Positive Signals (Ranked by Impact)
+
+### Creating Content That Scores High
+
+| Signal | How to Optimize |
+|--------|-----------------|
+| **Reply + Author Reply** | Engage back with every reply within 1 hour |
+| **Good Click** | Strong opening hook → valuable thread → clear CTA |
+| **Video 50% Completion** | 15-45 second videos, hook in first 3 seconds, captions |
+| **15s Tweet Dwell** | Long-form threads, data visualizations, deep analysis |
+| **Share** | Controversial takes, unique data, meme-worthy insights |
+| **Profile Click → Engage** | Consistent posting, clear bio, pinned quality tweet |
+| **Bookmark** | Tutorials, resources, reference material |
+
+### Real-Time Aggregation Windows
+
+The algorithm uses **real-time engagement signals** with specific lookback windows:
+
+```
+Signal                          Lookback Window
+────────────────────────────────────────────────
+Tweet Favorite (90D)            90 days
+Retweet (90D)                   90 days
+Reply (90D)                     90 days
+Good Tweet Click                Recent (real-time)
+Video Quality View              90 days
+Account Follow                  Infinite
+Repeated Profile Visit (14D)    14 days (min 2 visits)
+Repeated Profile Visit (90D)    90 days (min 6 visits)
+```
+
+---
+
+## Negative Signals (Avoid These)
+
+### The Algorithm Penalizes These Heavily
+
+| Signal | Weight Range | How It Hurts |
+|--------|--------------|--------------|
+| **Report Tweet** | -20,000 to 0 | Massive score reduction |
+| **Strong Negative Feedback** | -1,000 to 0 | Block/mute/report |
+| **Weak Negative Feedback** | -1,000 to 0 | "Not interested" clicks |
+| **Don't Like** | Negative | "See fewer" clicks |
+| **Unfollow** | Tracked | Reduces RealGraph score |
+| **Unfavorite** | Tracked | Counted in SimClusters |
+
+### Content That Triggers Negative Feedback
+
+From the visibility library and trust & safety models:
+
+1. **NSFW content** (when not expected)
+2. **Spam patterns** (repetitive posting, link farms)
+3. **Low-quality media** (pixelated, watermarked)
+4. **Engagement bait** without delivery
+5. **Misleading content** that doesn't match promise
+6. **Reply spam** to high-follower accounts
+
+---
+
+## The Four Pillars of Recommendation
+
+### 1. RealGraph (Relationship Scoring)
+
+**What it does**: Predicts likelihood you'll interact with another user
+
+**Key threshold**: Score ≥ 0.26 needed for recommendations
+
+**How it's calculated**:
+- Favorites exchanged
+- Retweets exchanged
+- Replies exchanged
+- Profile visits
+- Follow relationship
+- Address book data
+
+**Optimization**: Build genuine relationships with diverse, engaged accounts
+
+### 2. SimClusters (Community Detection)
+
+**What it does**: Places you and your content into ~145,000 communities
+
+**Key thresholds**:
+- Minimum 8 favorites to enter cluster
+- Score ≥ 0.3 to be retained
+- Cosine similarity ≥ 0.7 for high match
+
+**Half-life decay**: 8 hours (recent engagement matters most)
+
+**How tweets enter clusters**:
+```
+Tweet Embedding = Σ(InterestedIn vectors of users who favorited)
+```
+
+**Optimization**:
+- Get favorited by users in your target communities
+- Post about topics your audience cares about
+- Engage with content in your niche
+
+### 3. TwHIN (Knowledge Graph Embeddings)
+
+**What it does**: Dense vector representations for users and tweets
+
+**Models used**:
+- `TweetBasedTwHINRegularUpdateAll20221024`
+- `ConsumerBasedTwHINRegularUpdateAll20221024`
+- Collaborative filtering models for follows/engagement
+
+**How it works**:
+- Your user embedding = your engagement history
+- Tweet embedding = aggregated user interactions
+- ANN search finds similar content
+
+**Optimization**: Engage consistently with accounts similar to your target audience
+
+### 4. UTEG (User Tweet Entity Graph)
+
+**What it does**: Real-time "liked by" recommendations
+
+**Key constraints**:
+- Only keeps 24-48 hours of data
+- Max 800 out-of-network candidates
+- Min 1 favorited-by user
+
+**How it works**:
+- Traverses weighted follow graph
+- Finds tweets engaged by your network
+- Weights by influencer status
+
+**Optimization**:
+- Get liked by accounts with large, engaged followers
+- Timing matters (recency crucial)
+- Diverse engagement types help (likes + retweets + quotes)
+
+---
+
+## Actionable Optimization Strategies
+
+### Strategy 1: Maximize Reply Engagement
+
+The `ReplyEngagedByAuthor` signal is one of the highest-weighted:
+
+```
+1. Respond to EVERY reply within 60 minutes
+2. Ask follow-up questions in responses
+3. Create threads that invite discussion
+4. Use polls to generate reply activity
+5. End tweets with questions
+```
+
+### Strategy 2: Optimize for Dwell Time
+
+Both `TweetDetailDwell15s` and `ProfileDwell20s` are tracked:
+
+```
+1. Write substantive threads (5-10 tweets)
+2. Include data visualizations
+3. Use carousel images
+4. Create content worth re-reading
+5. Pin your best content for profile dwell
+```
+
+### Strategy 3: Video Completion Rate
+
+`VideoPlayback50` and `VideoQualityWatch` matter:
+
+```
+1. Hook in first 3 seconds
+2. Keep videos 15-45 seconds optimal
+3. Add captions (85% watch without sound)
+4. Front-load value
+5. Use native upload (not links)
+```
+
+### Strategy 4: Trigger Good Clicks
+
+`GoodClickV1` and `GoodClickV2` are high-value:
+
+```
+GoodClickV1: Click → View Detail → Favorite OR Reply
+GoodClickV2: Click → 2+ user actions taken
+
+1. Write compelling opening lines
+2. Promise value, deliver in thread
+3. Use curiosity gaps appropriately
+4. Include clear calls to action
+5. Make threads scannable with structure
+```
+
+### Strategy 5: Build RealGraph Strength
+
+Your relationship scores affect in-network visibility:
+
+```
+1. Engage authentically with 20-30 key accounts
+2. Reply to their content regularly
+3. Retweet with comments
+4. Exchange DM conversations
+5. Cross-platform interaction (if address book shared)
+```
+
+### Strategy 6: Enter the Right SimClusters
+
+Community placement affects out-of-network reach:
+
+```
+1. Identify which clusters your target audience is in
+2. Get favorited by users in those clusters
+3. Post consistently about cluster topics
+4. Engage with content in your niche
+5. Build relationships with cluster influencers
+```
+
+### Strategy 7: Timing for UTEG Visibility
+
+The 24-48 hour window is critical:
+
+```
+1. Post when your network is most active
+2. Engage with high-follower accounts before posting
+3. Get early engagement from influential followers
+4. Coordinate with friends for initial boost
+5. Reply to trending topics in your niche
+```
+
+### Strategy 8: Avoid Negative Signals
+
+One report can tank your score:
+
+```
+1. Don't post controversial content without value
+2. Avoid engagement bait that doesn't deliver
+3. Don't spam replies to celebrities
+4. Remove/delete poor-performing content
+5. Block trolls before they report you
+```
+
+---
+
+## API-Based Analysis Opportunities
+
+With your Twitter API access, you can perform several powerful analyses:
+
+### 1. Engagement Pattern Analysis
+
+```python
+# Analyze which of your tweets got the most engagement
+# Correlate with posting time, content type, length, etc.
+metrics_to_track = [
+    'like_count',
+    'retweet_count',
+    'reply_count',
+    'quote_count',
+    'impression_count',
+    'bookmark_count'
+]
+```
+
+### 2. Audience Cluster Analysis
+
+```python
+# Analyze who engages with your content
+# Map to potential SimCluster communities
+audience_analysis = {
+    'top_engagers': [...],
+    'their_interests': [...],
+    'shared_clusters': [...]
+}
+```
+
+### 3. Optimal Posting Time
+
+```python
+# Analyze when your audience is most active
+# When do your tweets get the fastest initial engagement?
+time_analysis = {
+    'best_hours': [...],
+    'best_days': [...],
+    'engagement_velocity': [...]
+}
+```
+
+### 4. Content Type Performance
+
+```python
+# Which content types perform best for your audience?
+content_types = [
+    'text_only',
+    'with_image',
+    'with_video',
+    'thread',
+    'poll',
+    'quote_tweet'
+]
+```
+
+### 5. Network Analysis
+
+```python
+# Who should you engage with more?
+# Which relationships drive the most value?
+network_analysis = {
+    'high_value_connections': [...],
+    'reciprocal_engagement': [...],
+    'potential_collaborators': [...]
+}
+```
+
+---
+
+## Technical Deep Dives
+
+### Key Files in the Algorithm
+
+| Component | Path | Purpose |
+|-----------|------|---------|
+| Main Pipeline | `home-mixer/server/.../ScoredTweetsRecommendationPipelineConfig.scala` | Orchestrates everything |
+| ML Scorer | `home-mixer/.../scorer/NaviModelScorer.scala` | Calculates weighted scores |
+| Engagement Signals | `home-mixer/.../scorer/PredictedScoreFeature.scala` | Defines 17 predictions |
+| Weight Params | `home-mixer/param/HomeGlobalParams.scala` | Model weight configuration |
+| Heuristic Rescoring | `home-mixer/.../scorer/HeuristicScorer.scala` | Post-ML adjustments |
+| SimClusters | `src/scala/com/twitter/simclusters_v2/` | Community embeddings |
+| RealGraph | `src/scala/com/twitter/interaction_graph/` | Relationship scoring |
+| UTEG | `src/scala/com/twitter/recos/user_tweet_entity_graph/` | Liked-by graph |
+| User Signals | `user-signal-service/` | Engagement tracking |
+
+### Heuristic Multipliers Applied
+
+```scala
+val rescorers = Seq(
+  RescoreOutOfNetwork,           // 0.75x for out-of-network
+  RescoreReplies,                // 0.75x for replies
+  RescoreMTLNormalization(...),  // Multi-task learning normalization
+  RescoreListwise(ContentExploration...),
+  RescoreListwise(DeepRetrieval...),
+  RescoreListwise(AuthorBased...),
+  RescoreListwise(ImpressedAuthorDecay...), // 0.5x decay for same author
+  RescoreListwise(MediaClusterDedup...),
+  RescoreFeedbackFatigue(...),   // Penalize recently shown
+  RescoreLiveContent             // Boost live/trending
+)
+
+// Score = baseScore × Π(rescorer multipliers)
+```
+
+### SimCluster Score Calculation
+
+```
+Tweet enters cluster when:
+1. Favorited by ≥8 users
+2. Score ≥ 0.3 in cluster
+3. Within 8-hour half-life window
+
+Score = Σ(engaged_user_cluster_weight × time_decay)
+Time decay = exp(-t / 8 hours)
+```
+
+---
+
+## Summary: The Growth Formula
+
+1. **Create content that earns "good clicks"** (click → dwell → engage)
+2. **Reply to every comment** to boost ReplyEngagedByAuthor signal
+3. **Post videos with 50%+ completion** optimization
+4. **Build genuine relationships** for RealGraph strength
+5. **Engage in your niche** to enter right SimClusters
+6. **Time posts for UTEG** visibility (first 24-48h crucial)
+7. **Avoid negative signals** at all costs
+8. **Optimize for dwell time** with substantive content
+
+---
+
+## Next Steps
+
+1. **Audit your recent content** against these signals
+2. **Set up API tracking** for pattern analysis
+3. **Identify your target SimClusters** through audience analysis
+4. **Build 20-30 key relationships** strategically
+5. **Create content calendar** optimized for these signals
+
+---
+
+*This analysis is based on Twitter's open-sourced algorithm repository as of the commit history in this codebase. Algorithm weights and parameters may be adjusted by Twitter without public disclosure.*