Open Source Code Breakdown · Commit a389166
An interactive, visual map of the «For You» feed: from thousands of source candidates through Phoenix scoring and safety checks to the final result set.
One feed request
A deterministic pipeline of 7 decisions, not a single opaque black box.
Real-time dynamic computation
30-Second Summary
Every feed request follows these four phases; exact live latency is not published in the repository.
Full Journey
Click any stage to explore its mechanics, interactive visual simulation, input/output data, and practical takeaways.
Stage 1 of 7
Before touching posts, the system inspects the viewer: recent actions, followed creators, active communities, language, and muted keywords.
The request combines recent engagement sequences with follows, blocks, mutes, locale, and other viewer context.
Real-world Example: If you spent the last 10 minutes watching machine learning demos and muted political keywords, those two signals immediately bias your search space.
Why it matters: That is why the exact same post receives completely different probability distributions for different people. There is no universal post score.
Snippet is pinned to commit a389166, ensuring permanent reproducibility.
Open pinned snippet on GitHubCandidate Retrieval
The system gathers candidates from distinct sources to balance familiar creators you follow with relevant discoveries.
Fresh posts from accounts you actively follow. Posts you already saw are filtered out at query time.
Configuration limits for candidate sources and pool sizes.
Open source candidate limits on GitHubTransforms your recent reading history into an embedding vector and retrieves nearest-neighbor posts globally.
Configuration limits for candidate sources and pool sizes.
Open source candidate limits on GitHubFinds posts that are trending within your shared interest clusters. Public cosine similarity threshold is >0.5.
Configuration limits for candidate sources and pool sizes.
Open source candidate limits on GitHubCounts are code-level ceilings and defaults, not a promise that every request fills each stage.
Interactive Simulation
Test how predicted engagement probabilities and negative penalties combine into the final ranking score.
Simulated Base Score
0.455Production formula combines calibrated log-odds and dynamic experiment offsets.
Open ranking scorer formula on GitHubFrequent Action · Low Weight
12% Like chance × 0.5 = +0.060Common action, provides steady baseline support.Rare Action · High Weight
0.4% Copy Link chance × 20 = +0.080Low probability, but delivers 33% more score impact than the like!Which post from this author is it in the current candidate pool?
Public Defaults
In the open source codebase, like is the baseline unit (weight = 0.5). Compare how other interactions scale relative to a like.
Phoenix first predicts probabilities [0.0...1.0]. Then public weights turn them into mathematical score contributions.
What boosts score when predicted with equal probability
| Predicted Action | Scale | Weight | vs Like |
|---|---|---|---|
| Copy linkHighest organic amplification signal | +20 | 40× | |
| Reply to mutual follow5.0 base + 15.0 conversation boost for mutual connections | +20 | 40× | |
| ReplyActive conversation starter | +5 | 10× | |
| Share via Direct Message (DM)Private peer recommendation | +5 | 10× | |
| Quote tweetContextual commentary | +5 | 10× | |
| Follow authorDirect creator affinity signal | +4 | 8× | |
| Share (native share sheet) | +2 | 4× | |
| Repost (Retweet) | +1 | 2× | |
| Like (Favorite)baseline | +0.5 | 1× | |
| Post clickTapping to expand tweet detail | +0.4 | 0.8× | |
| URL click (External link)Navigating to linked website | +0.2 | 0.4× | |
| Photo expansion | +0.05 | 0.1× | |
| Video playback started | +0.05 | 0.1× | |
| Video quality view (>5s) | +0.05 | 0.1× | |
| Quote tweet click | +0.05 | 0.1× | |
| Unexplored post boostExploration bonus enabled by default in-network | +0.02 | 0.04× | |
| Continuous reading dwell timeScore per second of sustained reading dwell | +0.004 | 0.008× |
What drastically reduces score when predicted
| Negative Action | Log Scale | Penalty Weight |
|---|---|---|
| Report postSevere violation signal (468× the like coefficient at equal predicted probability) | -234 | |
| Mute authorViewer explicitly silences author (about 118× the like coefficient) | -58.8 | |
| «Not interested in this»Explicit negative feedback in feed (about 86× the like coefficient) | -43.2 | |
| Block authorHard social cut (about 62× the like coefficient) | -31.2 | |
| Did not linger (Immediate bounce)User immediately scrolled away without reading | -0.02 |
Logarithmic scale: otherwise a -0.02 bounce penalty would be invisible next to a -234 report penalty.
Compare public coefficient magnitudes using like weight (+0.5) as the baseline. This is not a reach forecast.
Public default scoring parameters and penalty definitions.
Open parameters file on GitHubPost-Scoring Adjustments
Recommendations optimize for the entire feed experience: the algorithm penalizes repetitive posting, applies network distance thresholds, gives trial slots to new creators, and balances topics.
Multiplier = 0.75 × 0.5^(n-1) + 0.25
This is a continuous decay curve, not a hard ban. Post #1 keeps 100%, Post #2 gets 62.5%, Post #3 gets 43.75%, Post #4 gets 34.38%, approaching a floor of 25%.
Posts from creators you do not follow receive a default 0.75× multiplier, prioritizing relevant content from your existing social graph.
When all experiment, rank, freshness, impression, and safety gates pass, the scorer can raise one best eligible original post toward the score at zero-based slot 15. This is not guaranteed placement.
Determinantal Point Processes (DPP) mathematically balances individual relevance against mutual semantic similarity.
Myths vs Reality
Separating social media folklore from verified open source code facts.
«Likes are the primary goal for reach»
«Bookmarks are ignored by the algorithm because there's no weight in the formula»
«External links will kill your reach»
«Posting 10 times a day multiplies your chances»
«Replying to viral accounts exposes you to their audience»
«Safety filters just dock a few score points»
«Hashtag stuffing guarantees topic categorization»
Practical growth system
Follower count is not a public ranking weight. Use these ranges to locate your current constraint, then move on when you have the proof for the next stage. Each plan is designed for your next ten standalone posts—not as a promise of reach.
The public defaults set both the author follower cap and the post impression threshold at 1,000. An eligible standalone original or quote post can be considered by the cold-start scorer. The 0–100, 101–500, and 501–1,000 ranges are planning bands, not extra ranking thresholds.
Eligibility only puts one qualifying candidate in position for a possible test. It still needs enough initial rank, the right experiment and corpus conditions, freshness, visibility, and a strong viewer-specific prediction.
A new reader cannot tell in a few seconds who the account is for or why they should come back.
One clear promise, three repeatable subject areas, and the first substantive responses from intended readers.
Posting volume, broad topic coverage, and viral benchmarks.
Audience → problem → useful outcome in one sentence.
Ten standalone posts built around real questions.
Useful replies that bring the right people back.
Use: “I help [audience] solve [problem] through [kind of value].” Make the bio, pinned post, and recent feed support the same sentence.
A follow decision is easier when the post and profile promise the same result. This is an editorial strategy, not a ranking coefficient.
Collect the exact language from conversations, search, comments, and people in the niche. Each question should be specific enough to become a standalone post.
A concrete question anchors the post in an observed need instead of a guessed topic.
Use three shapes: error teardown, evidence-led observation, and short instruction. Give every post its own context, claim, and payoff.
The cold-start path excludes replies and reposts. A standalone original or substantive quote post has the clearer route to an unfamiliar viewer.
After publishing, find several discussions around the same problem and add a fact, example, or clarifying question. Save the language that earns substantive replies.
Replies can build relationships and reveal audience language, but they do not automatically expose you to a large author’s audience.
For each of the ten posts, record subject, shape, substantive replies, shares, profile visits, and follows. Decide only after several attempts.
The public scorer combines many predicted actions. One like spike or impression spike does not prove a durable format.
Choose three questions from one audience and one shared problem.
→Answer them completely in three different post shapes.
→Record who responded and which action repeated.
The published defaults set the author follower cap and the eligible post impression threshold to 1,000, with cold-start enabled by default in this snapshot.
Open-source codeSmall-author cold-start pathEligibility excludes replies and reposts, checks the follower cap and impression threshold, then promotes only the best eligible candidate when the remaining gates pass.
Open-source codeReplies serve relationships, not broad OON discoveryThe For You pipeline removes out-of-network replies and reposts before scoring. Original standalone posts have a clearer route to non-followers.
Official X guidanceConversation builds communityX recommends useful replies, questions, listening, and a recognizable human voice for relationship-building—not automated reply volume.
Official X guidanceTest concise, conversational creativeX advises concise copy, restrained hashtags, clear calls to action where useful, accessible media, and repeated testing against real response.
Practical Implications
Not a get-rich-quick blueprint. These are mechanical consequences derived directly from the codebase logic.
The author's 2nd post in the current candidate pool receives a 0.625× multiplier, 3rd gets 0.4375×, 4th gets 0.34375×.
Code snippet verifying this architectural behavior.
Open source reference on GitHubReplies and reposts from accounts outside the viewer's network are filtered out before scoring in this pathway.
Code snippet verifying this architectural behavior.
Open source reference on GitHubPhoenix explicitly predicts negative reactions, and public penalty weights (-234 for reports, -58.8 for mutes) overpower hundreds of positive likes.
Code snippet verifying this architectural behavior.
Open source reference on GitHubThe primary candidate pipeline for For You hard-filters all posts older than 48 hours.
Code snippet verifying this architectural behavior.
Open source reference on GitHubA click carries a modest weight of 0.4. Immediate bounces (-0.02) and lack of follow-through engagement yield weak overall scores.
Code snippet verifying this architectural behavior.
Open source reference on GitHubDownstream from the ML model, DPP diversity, safety policies, dynamic A/B experiments, and anti-gaming heuristics all take effect.
Code snippet verifying this architectural behavior.
Open source reference on GitHubExecutive Summary
The core architectural rules of the open source recommendation engine distilled into 5 key principles.
The exact same post produces completely different predictions for different viewers based on their active interest vectors.
Candidate eligibility strictly expires after 48 hours in the primary recommendation pipeline.
A single link copy (+20.0) contributes as much score as 40 individual likes (+0.5).
Out-of-network replies are filtered out, while original standalone posts remain eligible to reach global audiences.
Candidates compete on weighted ML scores first, but independent visibility rules can still drop posts before rendering.
Scope & Transparency
Understanding the boundaries between the public repository snapshot and live production servers.
We can understand the exact mechanics, pipeline stages, and public defaults. We cannot promise «do X and get Y impressions»: the model is deeply personalized, parameters are subject to dynamic experimentation, and anti-abuse safeguards remain closed.
Mastery Check
Test your comprehension with 10 challenging, real-world questions grounded directly in the open-source codebase. Review your results with in-depth technical breakdowns.
README.mdExact Pipeline Execution Orderhome-mixer/candidate_pipeline/phoenix_candidate_pipeline.rsPublic Parameters & Weightshome-mixer/params/param.rsRanking Scorer Formulahome-mixer/scorers/ranking_scorer.rsNew Author Cold Start Injectionhome-mixer/scorers/author_cold_start.rsDPP Semantic Topic Diversityvm-ranker/dpp.rsVisibility Rules & Safety Registryvisibility-filtering/rules/registry.rsUnpublished Production BoundariesREADME.md