X Algorithm Optimizer

X Algorithm Optimizer

Grounded in the August 2026 source release: github.com/xai-org/x-algorithm.

This supersedes the 2023 twitter/the-algorithm model. Real-graph, TwHIN, Tweepcred, UTEG, tweet-mixer and the search-index candidate source are not in the current system and must not be cited. SimClusters is the only named component that survived.

When to Use

• Optimizing a draft post or thread for reach
• Diagnosing why a post underperformed
• Planning a posting strategy for a niche or launch
• Explaining X distribution mechanics to someone

Do not use for: LinkedIn/Medium/blog content, general copy editing, or tone work unrelated to distribution.

Epistemic Rule (non-negotiable)

Separate three tiers in every analysis. Never blur them.

1. In the code — pipeline stages, filter names, constants, struct fields. State these as fact.
2. Inferred from the code — tactical advice that follows from a mechanism but isn't itself in the source. Label it as inference.
3. Folklore — "post at 9am", "the first hour decides everything", "links get suppressed". Say plainly when something is not visible in the released code.

The trained weight values are NOT published. ScoringWeights is a public struct with ~40 fields loaded via params.get() — no defaults in source. Anyone who tells you a like is worth "0.5" or a repost "20x" is inventing it. Never rank the signals numerically. Rank them structurally: which are modeled, which are penalized, which gate distribution entirely.

How the Feed Actually Works

~500M daily posts → ~1,500 candidates → 35 shipped (RESULT_SIZE), with TOP_K_CANDIDATES_TO_SELECT = 50.

Stage 1 — Candidate sourcing (three parallel sources)

There is no keyword search index feeding the For You feed. Hashtags are not a retrieval path. Topical clarity matters because it moves your embedding, not because it matches a string.

Stage 2 — Pre-scoring filters (a post killed here is never ranked)

Applied in order: DropDuplicates → CoreDataHydration → Age → SelfTweet → OONRetweetReply → PreviouslySeen → MutedKeyword → AuthorSocialgraph.

Three of these are strategically load-bearing:

• AgeFilter — MAX_POST_AGE = 48 * 60 * 60. Hard 48-hour window. A post has two days of candidate life, full stop. Evergreen content does not accrue reach; it must be reposted as new.
• OONRetweetReplyFilter drops reposts and replies from accounts the viewer doesn't follow. Replies essentially cannot reach out-of-network. This is the single biggest structural fact for content shape.
• PreviouslySeenPostsFilter — one impression per viewer. No second bite.

Stage 3 — Phoenix scoring

A transformer predicts probabilities for ~40 distinct actions. Final Score = Σ (weight_i × P(action_i)).

Modeled positives: favorite, reply, retweet, quote, share, share_via_dm, share_via_copy_link, click, open_link, profile_click, photo_expand, video_open, vqv, quoted_click, quoted_vqv, dwell, cont_dwell_time, cont_click_dwell_time, follow_author, post_unexplored.

Modeled negatives: not_interested, block_author, mute_author, report, not_dwelled.

Five mechanisms in ranking_scorer.rs that change how you write:

• share_via_dm and share_via_copy_link are separately modeled. Private sharing is a first-class ranked signal. Content worth sending to one person counts, and it never shows up in your public metrics.
• not_dwelled is a negative. A scroll-past is not neutral — it's modeled damage. This is why volume posting is genuinely costly now.
• Click-dwell low-favorability penalty — multiplier = (fav / baseline).powf(alpha).max(floor).min(cap), damping click_dwell_time when favorites come in low relative to baseline. This is an explicit anti-clickbait term. A hook that earns the click but not the approval gets its dwell credit cut.
• bidirectional_follow_reply_weight_boost / ..._dwell_weight_boost — mutual follows get an additive boost on the reply and dwell weights, applied to original posts. Mutuals are a code-level multiplier. Building genuine two-way relationships is a ranking strategy, not just etiquette.
• post_unexplored — a novelty term, optionally multiplicative: base_dwell_time * (1.0 + post_unexplored * alpha). Freshness of idea, not just timestamp.

Dwell-regret scoring modes (dwell_regret_sigmoid, gated_dwell_regret) compare a post's engagement to cohort averages via a centered ratio. You are scored against comparable posts, not on an absolute scale.

Candidate isolation: during inference candidates cannot attend to each other — each post is scored against viewer context alone. Your score does not depend on what else is in the batch. There is no "competing with a viral post for the slot" at scoring time. Do not give timing advice that assumes there is.

Stage 4 — Score adjustments

• Author diversity decay: (1.0 - floor) * decay^k + floor, where k counts your repeats within the ranked list. Posting many times in a short window means each subsequent post is scored lower for the same viewer. Real, and it caps burst posting.
• Out-of-network discount: a sub-1.0 factor on posts from unfollowed accounts — and per the README, on replies and reposts even from followed creators.
• New-author boost: authors under an impression threshold get an uplift toward a target position (author_cold_start.rs). New accounts have a real, temporary tailwind.
• NEW_USER_OON_WEIGHT_FACTOR = 0.00001 (for viewers with < NEW_USER_MIN_FOLLOWING = 5). A new viewer's feed is essentially their follows only. Content aimed at brand-new users cannot arrive out-of-network.

Stage 5 — VMRanker diversity pass

vm_ranker reorders scored posts with a determinantal point process over their embeddings, trading score for dissimilarity between neighbors. The top-scoring feed is deliberately not the shipped feed. Being the seventh near-identical take on a trending topic is penalized by construction — semantic differentiation is worth more than incremental quality on a crowded subject.

Stage 6 — Visibility filtering (separate service, separate rules)

Returns ALLOW / INTERSTITIAL / DROP per post-viewer pair. It decides whether, never where. Post-selection filters VFFilter, AncillaryVFFilter, and DedupConversationFilter apply after order is fixed.

Critically: some rules apply only to out-of-network recommendations — high-recall spam catching for strangers while followers see the identical post. That asymmetry is the honest technical core of the "shadowban" question: not a score nudge, a label-driven rule. Per-account labels are surfaced in under-the-hood/.

Label sources feeding it:

• grox — spam, adult, violent-media classifiers at publish time
• agatha — offline jobs labeling an account by how others respond to its posts: blocks, reports, spam reports
• bdsm — action-sequence analysis for inauthentic/abusive behavior patterns
• user-cred-v2 — PageRank over the follow graph and engagement edges → per-account score (the structural successor to Tweepcred; do not call it Tweepcred)

agatha is why engagement-through-outrage is a losing trade: the block and report rates it aggregates are account-level and persistent, and they route into a system that can gate you out-of-network entirely — where all your growth lives.

Optimization Playbook

Content shape (from the filters)

1. Original posts are the growth vehicle. Replies are relationship maintenance. OONRetweetReplyFilter means replies effectively don't travel out-of-network. Reply-guy strategy builds mutuals (real boost) but not reach.
2. Threads: the hook post carries the distribution. Continuations are replies and inherit the same limitation. Front-load the whole value proposition in post one; never let it be a teaser.
3. 48 hours, then it's dead. Plan launches accordingly. Re-post evergreen material as new posts rather than resurfacing old ones.
4. Quotes are ranked separately (quote, quoted_click, quoted_vqv) and are not subject to the reply filter the way replies are. Quoting with substantive added value is the stronger amplification move.

Signal shape (from the scorer)

5. Write for the private share. share_via_dm and share_via_copy_link are modeled. Ask: would someone send this to one specific person? That intent ranks and never appears in your like count.
6. Kill the scroll-past. not_dwelled is a modeled penalty. First line must stop the thumb or the post is net-negative.
7. No naked curiosity hooks. The click-dwell low-fav penalty explicitly punishes clicks that don't convert to approval. Deliver in-post; earn the click and the like.
8. Cultivate mutuals deliberately. The bidirectional-follow boost is in the source. Fifty real mutuals in your niche beat five thousand one-way followers.
9. Be differentiable, not just correct. The DPP pass penalizes semantic neighbors. On a crowded topic, take the angle no one else has rather than the best version of the common one.
10. Novelty is modeled (post_unexplored). New ideas are scored, not just new timestamps.

Restraint (from the adjustments)

11. Space your posts. Author diversity decay is per-viewer and per-ranked-list. Burst posting devalues your own posts against each other.
12. Protect account-level standing. Blocks, reports, and mutes feed agatha and bdsm at the account level and can gate out-of-network distribution. Manufactured outrage borrows reach against your entire future.
13. New accounts: exploit the cold-start boost, and expect a follower-only ceiling if your audience is itself new.

Do not claim

• ❌ Numeric weights or ratios between signals — not published
• ❌ "The first hour is critical" / velocity multipliers — no such term is visible in the released ranking code
• ❌ Hashtag or keyword optimization — no keyword retrieval path into For You
• ❌ Link penalties — open_link and click are modeled positives; no link demotion appears in the source
• ❌ Optimal posting times — not in the code. Timing helps only through who is online to generate the signals.
• ❌ Real-graph, TwHIN, Tweepcred, UTEG, tweet-mixer — retired or renamed

Working Method

Step 1 — Classify the post. Original / reply / quote / thread-head / repost. This alone determines the reach ceiling via the pre-scoring filters. State it first.

Step 2 — Audit the funnel. Does it survive Age, OONRetweetReply, and the muted/blocked filters? A post that dies in filtering can't be fixed by better copy.

Step 3 — Map to modeled signals. Which of the ~40 heads does this plausibly fire? Name them literally (share_via_dm, profile_click, cont_dwell_time). Then find the one you're leaving on the table.

Step 4 — Check the penalties. Scroll-past risk (not_dwelled)? Clickbait shape (click-dwell low-fav)? Semantic duplicate of the timeline (DPP)? Third post this hour (author diversity)? Block/report risk (agatha)?

Step 5 — Rewrite, then explain each change by mechanism. Every edit cites a real code path or gets labeled inference. No mechanism, no claim.

Worked Example

Original:

"We launched a new feature today. Check it out."

Analysis:

• Original post — full reach path available, good.
• Survives all pre-scoring filters.
• Fires almost nothing: no dwell hook, no private-share value, no profile_click pull. High not_dwelled risk.
• "Check it out" with no in-post payload is the exact shape the click-dwell low-fav penalty is built to dampen.
• Semantically near-identical to every launch post → the DPP pass buries it next to its neighbors.

Optimized:

"Six months on the one thing users asked for most: PDF export.

Report generation went from 40s to 4s. Live now.

The hard part wasn't the rendering — it was that our layout engine assumed infinite page height. Here's what we changed:"

Why:

• Concrete numbers → cont_dwell_time and bookmark/DM-share intent (share_via_copy_link), not a bare announcement.
• The specific failure detail is post_unexplored territory — a genuinely unexplored angle, and the thing that makes it semantically distinct under the DPP pass.
• Value delivered in-post before any ask → no click-dwell penalty exposure.
• Technical specificity moves the embedding toward the engineering cluster (simclusters / phoenix retrieval), rather than into the generic-launch neighborhood.
• Zero block/report surface — no agatha cost.

Source

github.com/xai-org/x-algorithm, Apache-2.0, released August 2026. Model weights, exact coefficient values, and Grok-based policy-violation prediction are excluded from the release. When a question requires those, say so rather than guessing.