User Pattern Scenarios & Ranking Implications
Concrete mathematical calculations demonstrating how different creator posting behaviors and audience engagement profiles trigger the algorithm's utility engine.
Behavior: Creator shares an authoritative technical infographic or architecture diagram. Users inspect the image for 30 seconds and copy the link to share into company Slack/Discord channels.
Ranking Implication: High-intent off-platform advocacy triggers a massive scoring surge. The post easily surpasses candidate thresholds in both In-Network and Out-of-Network feeds, expanding into broader SimClusters topic graphs.
Behavior: Creator poses a nuanced engineering question and actively engages with reciprocal followers in multi-turn comment threads.
Ranking Implication: Conversational threads with mutual followers score higher than 1,000 passive likes (1,000 × 0.5 = 500.0). The Phoenix Scorer's conversation detection boosts root post distribution to connected network clusters.
Behavior: Creator publishes 5 standalone tweets in 20 minutes instead of a single unified thread.
• Post 1 (k=0): 100.0 × 1.000 = 100.0
• Post 2 (k=1): 100.0 × 0.625 = 62.5 (-37.5%)
• Post 3 (k=2): 100.0 × 0.4375 = 43.75 (-56.25%)
• Post 4 (k=3): 100.0 × 0.3438 = 34.38 (-65.62%)
• Post 5 (k=4): 100.0 × 0.2969 = 29.69 (-70.31%)
Ranking Implication: Flooding standalone posts within a short window severely penalizes subsequent tweets. Top-K selection filters out posts 2–5. Multi-point discussions must be formatted as linear reply threads or spaced by ≥2 hours.
Behavior: Account publishes polarizing outrage bait that generates quick clicks but offends viewers.
Negative Penalties = (2 Reports × -234.0) + (3 Mutes × -58.8) + (1 Block × -31.2) = -468 - 176.4 - 31.2 = -675.6
Net Utility Score = 300.0 - 675.6 = -375.6
Ranking Implication: Net-Negative Normalization forces this post below ALL positive content on the entire platform. Additionally, account-level Scarecrow trust vectors degrade, triggering downstream DROP visibility verdicts.
Behavior Profile: Creator shares research via a bare external URL preview card vs. sharing an inspectable infographic summary with the external destination URL in the first thread reply.
Algorithmic Verdict: While external links are not assigned an explicit negative weight (openLink = +0.2), they prematurely terminate in-app dwell. Placing outbound links in the first reply thread captures massive in-app dwell (+0.004) and mutual conversation bonuses (+15.0) on the parent post while still driving outbound traffic.
Tactical HowTo Guides by Behavioral Pattern
Actionable procedural workflows tailored to each algorithmic dynamic.
HowTo: Engineer Posts for Copy-Link Maximization (+20.0)
1. Synthesize High-Density Frameworks: Condense complex systems into clean cheat sheets or comparison matrices.
2. Attach High-Contrast Infographics: Embed visual assets that demand >20 seconds of inspection to build Continuous Dwell (+0.004).
3. Write Bookmarkable Openers: Clearly state the reference utility in the first line so readers copy the link for external work channels.
HowTo: Foster Mutual-Follow Conversations (+15.0 Bonus)
1. Seed Open-Ended Domain Debates: Conclude posts with specific questions that prompt domain peers to share experiences.
2. Engage Within 60 Minutes: Reply promptly to mutual followers to create multi-turn chains that compound the +15.0 bonus.
3. Cultivate Bilateral Follow Graph Edges: Follow back active domain creators to establish reciprocal graph ties.
HowTo: Pace Publishing to Circumvent Author Decay
1. Enforce 2–4 Hour Spacing: Allow candidate generation caches to refresh before publishing another standalone tweet.
2. Consolidate into Linear Threads: Post secondary points as thread replies to accumulate dwell on the parent entity without author decay.
3. Disable Social Blast Schedulers: Prevent automated tools from dumping simultaneous standalone updates.
HowTo: Audit Content to Avoid Net-Negative Penalties
1. Verify Technical Claims: Provide documentation or commit links to prevent spam/misleading reports (-234.0).
2. Eliminate Polarizing Outrage: Avoid ragebait; 1 mute (-58.8) wipes out 118 positive likes.
3. Target Domain Niches: Ensure precise terminology so SimClusters doesn't expose your post to disinterested users who click 'Not Interested' (-43.2).
HowTo: Share External Links Without Incurring Dwell Reach Dampening
1. Synthesize Core Value in Primary Post: Deliver key insights, code snippets, or takeaways directly in the main post so reading delivers complete standalone value without requiring an off-platform click.
2. Place Destination URLs in First Reply: Keep the parent post free of outbound URL preview cards by dropping your article, repository, or product link into the immediate self-reply comment.
3. Pair with Inspectable Infographics: If a root URL is required, attach high-density visual assets demanding >20s inspection to accumulate Continuous Dwell (+0.004) and avoid the fast scroll penalty (-0.02).
How the Feed is Built: The 10-Stage Pipeline
Every timeline query executes an asynchronous distributed pipeline across candidate generation, ML scoring, diversity decay, and downstream visibility screening.
User opens app or scrolls past viewport boundary; triggers authenticated request with client capabilities.
Fetches recent engagements, follow graph edges, block/mute lists, negative keywords, recently seen post bloom filters, and real-time topic embeddings.
In-network (Thunder): Recent tweets from accounts the user follows.
Out-of-network (Phoenix + SimClusters): Semantic embedding clusters and vector search finding candidate posts from non-followed accounts.
Attaches full tweet text, media metadata (aspect ratios, video length), author reputation vectors, language tags, and live social engagement counters.
Drops duplicates, posts older than 48 hours, author self-posts, muted/blocked authors, posts in the user's seen-cache, and inaccessible subscription paywalls.
Deep Transformer model evaluates candidate posts in complete candidate isolation, predicting fine-grained per-action probabilities P(action).
Computes dot product utility scores, applies author diversity decay, out-of-network discount (0.75), cold-start lifts, and net-negative normalization boundaries.
Ranks and selects the top-K highest scoring posts across thousands of scored candidates into a structured timeline queue.
Downstream gate evaluating Scarecrow/Botmaker classifiers, credibility scores, suspension states, subscriber rules, and viewer legal settings -> produces ALLOW, INTERSTITIAL, or DROP.
Hydrated and verified feed delivered and rendered on the client timeline.
Scoring Mechanism & Signal Weights
The Phoenix Transformer Scorer predicts discrete action probabilities for each candidate post, which the RankingScorer combines into a total expected utility score.
| Signal Category | Action / Behavior | Weight | Strategic Intention & Dynamic |
|---|---|---|---|
| Advocacy (Positive) | Copy-Link Share | +20.0 | Highest positive signal. Rewards external virality and high-intent off-platform sharing. |
| Community (Positive) | Mutual-Follow Reply Bonus | +15.0 | Bonus awarded when mutual followers converse (+5.0 base + 15.0 bonus = +20.0 total). |
| Conversation (Positive) | Reply to Post | +5.0 | Promotes active on-platform discussion and multi-turn threads. |
| Discourse (Positive) | Quote Tweet | +5.0 | Incentivizes commentary, analysis, and perspective-sharing. |
| Private Sharing (Positive) | DM Share | +5.0 | Rewards 1-on-1 private peer recommendations. |
| Affinity (Positive) | Follow Author | +4.0 | Explicit signal that post motivated creating a new graph edge. |
| Amplification (Positive) | Native Share Sheet | +2.0 | Standard OS share dialog trigger. |
| Broadcast (Positive) | Repost (Retweet) | +1.0 | Standard broadcast redistribution (weighted 20x lower than copy-link). |
| Passive (Positive) | Favorite (Like) | +0.5 | Low-friction affirmation (weighted 40x lower than copy-link). |
| Navigation (Positive) | Post Click / Link Open | +0.4 / +0.2 | Detail expansion and external URL clicks. |
| Attention (Positive) | Continuous Dwell (per unit time) | +0.004 | Rewards sustained active reading/watching time in viewport. |
| Penalty (Negative) | Not Dwelled (Fast Scroll-by) | -0.02 | Minor penalty for scrolling past without stopping. |
| Severe Penalty (Negative) | Block Author | -31.2 | Heavy negative reaction indicating strong creator aversion. |
| Explicit Penalty (Negative) | Not Interested | -43.2 | Explicit user feedback dismissing author or topic. |
| Severe Penalty (Negative) | Mute Author | -58.8 | Severe penalty; wipes out the value of ~118 likes. |
| Catastrophic Penalty (Negative) | Report Post | -234.0 | Catastrophic penalty. Equivalent to 468 likes or 11 copy-links. Crushes reach immediately. |
Copy-Link Share +20.0
Highest positive signal. Indicates external off-platform virality and extreme user advocacy.
Mutual-Follow Reply Bonus +15.0
Accelerates reciprocal thread conversations (+5.0 base + 15.0 bonus = +20.0 total).
Replies, Quotes & DM Shares +5.0 each
Promotes active discussion, substantive commentary, and private peer recommendations.
Follow Author +4.0
Explicit social graph expansion motivated by post quality.
Report Post -234.0
Catastrophic penalty wiping out 468 likes or 11 copy-links. Triggers Net-Negative Normalization.
Mutes (-58.8), Not Interested (-43.2), Blocks (-31.2)
Massive asymmetric negative penalties ensuring polarizing content is algorithmically suppressed.
Post-Scoring Adjustments
Before candidates are sent to the VMRanker top-K selection stage, four critical balancing transformations are applied to prevent monopoly, balance discovery, and incentivize quality.
Author Diversity / Repeated-Author Decay
Default parameters: decay = 0.5, floor = 0.25.
When an author posts multiple candidate tweets in a single batch:
- Post 1: 100% full score
- Post 2: 62.5% score
- Post 3: 43.75% score
- Post 4+: 25% floor
Out-of-Network (OON) Discount
Recommended candidate posts from accounts the user does not follow are automatically discounted by 25%.
Takeaway: In-network followers provide a 33% structural advantage over out-of-network algorithmic discovery.
Cold-Start Exploration Boosts
New or low-impression creators receive temporary algorithmic lift into candidate pools to test baseline engagement against representative user clusters, solving the cold-start distribution problem.
Dwell-Regret Modeling
Post-hoc satisfaction filter that penalizes clickbait titles and sensationalist thumbnails that achieve high initial click-throughs but lead to abrupt bounces or post-dwell blocks.
Visibility Filtering (The Independent Gate)
Visibility filtering operates strictly downstream of ranking. It evaluates trust, safety, and policy signals to produce one of three definitive verdicts.
Show Normally
Post passes all credibility, safety, and regional compliance checks. Displayed prominently in the For You feed.
Warning Overlay
Post contains sensitive media or disputed claims. Displayed behind a click-through warning screen.
Do Not Show
Violates safety, copyright, block graphs, or botmaker rules. Purged completely regardless of its ranking score.
Platform Intentions & Game-Theoretic Optimizations
Analyzing why specific weights were engineered into the August 2026 configuration release.
Why Copy-Link is +20.0
Copying a link represents off-platform advocacy (sharing to Slack, WhatsApp, iMessage, email). It acts as an organic user-acquisition vector for X while signaling peak content utility.
Why Reports are -234.0
Asymmetric penalty design destroys engagement baiting. If a post gets 100 likes (+50) but triggers 1 report (-234), its net score is -184, triggering Net-Negative Normalization to bury the post.
Why Mutual Replies are +15.0
Incentivizes real community formation and bilateral networking over broadcast noise. Conversations between mutuals generate the stickiest retention on the platform.
Technical Glossary & Architecture Terms
Core concepts and terminology governing X's open-source recommendation engine.
Decoupled Architecture
The intentional separation of mathematical ranking utility from safety/compliance visibility filtering.
Phoenix Scorer
Deep Transformer neural network scoring candidate tweets independently in candidate isolation.
SimClusters
Community detection algorithm grouping users and posts into latent semantic topic interest spaces.
Repeated-Author Decay
Exponential mathematical curve dampening successive posts from the same creator in a single timeline assembly.
Net-Negative Normalization
Boundary sorting rule forcing any post with an overall negative score strictly below all positive candidates.
Continuous Dwell
Continuous active viewport dwell tracking (+0.004) rewarding long-form reading and inspection.
Frequently Asked Questions
Answers to key questions regarding algorithm mechanics, penalties, and feed assembly.
Why are ranking and visibility decoupled into separate systems?
Why does Copy-Link Share have an enormous +20.0 weight compared to Favorite at +0.5?
How does the algorithm handle negative user feedback like Mutes and Reports?
What is Repeated-Author Decay and how does it affect posting cadence?
How do In-Network posts compare against Out-of-Network (OON) recommendations?
Does the algorithm punish external links in posts?
Direct vs. Indirect Penalty: The algorithm does not apply an explicit negative coefficient to external URLs (openLink is weighted at +0.2). However, external links inflict a massive indirect penalty by bouncing users off-platform, immediately terminating in-app viewport Continuous Dwell (+0.004) and preventing on-platform Mutual-Follow Reply (+15.0) cascades.
The Optimal Creator Strategy: Creators who post self-contained infographics and summarize key takeaways in the main post, while placing outbound links in the first reply thread, consistently achieve 20x–40x higher algorithmic reach compared to posting bare external URL preview cards.
Explore Knowledge Graph Interactively Click to Open ▼
SPARQL Query Recipes & Workbench Click to Open ▼
1. Query All Signal Weights (Positive vs Negative)
PREFIX schema: <http://schema.org/> PREFIX : <https://x.com/IntuitMachine/status/2088198582880854475#> SELECT ?signal ?name ?value ?description WHERE { ?signal a schema:PropertyValue ; schema:name ?name ; schema:value ?value ; schema:description ?description . } ORDER BY DESC(?value)
2. Query 10-Stage Feed Generation Pipeline
PREFIX schema: <http://schema.org/> SELECT ?position ?stepName ?stepDescription WHERE { ?step a schema:HowToStep ; schema:position ?position ; schema:name ?stepName ; schema:description ?stepDescription . } ORDER BY ?position
3. Query Behavioral Scenarios
PREFIX schema: <http://schema.org/> PREFIX rdfs: <http://www.w3.org/2000/01/rdf-schema#> SELECT ?scenario ?name ?description ?howto WHERE { ?scenario a schema:Action ; schema:name ?name ; schema:description ?description ; rdfs:seeAlso ?howto . }