{
  "_meta": {"verified": "2026-07-31", "note": "Each myth verdict cites the 2026 repo (xai-org/x-algorithm @ 0bfc279) or the frozen 2023 repo (twitter/the-algorithm @ c54bec0). 'Absent from open code' ≠ proven absent from production — withheld crates (params, visibility filtering, Grox prompts) could hide behavior. We say so every time; that's the difference between this site and the folklore."},
  "myths": [
    {
      "slug": "one-reply-equals-150-likes",
      "claim": "\"One reply is worth 150 likes\"",
      "verdict": "OUTDATED — 2023 numbers presented as current",
      "explain": "The 13.5-vs-0.5 reply/like ratio (27×, sometimes misquoted as 150×) comes from the 2023 twitter/the-algorithm heavy-ranker weights. The 2026 stack keeps separate reply and favorite heads, but every 2026 weight is withheld in an unpublished params module. Articles quoting exact 2026 ratios are recycling 2023 data.",
      "code": "2023: the-algorithm-ml heavy ranker weights. 2026: home-mixer/scorers/ranking_scorer.rs reads all weights from crate::params — not in the repo."
    },
    {
      "slug": "tweepcred",
      "claim": "\"Your TweepCred / PageRank score caps your reach\"",
      "verdict": "GONE from the 2026 open code",
      "explain": "TweepCred (the PageRank-style author score with the famous follower/following penalty) exists only in the 2023 repo. The 2026 ranker consumes learned embeddings — 'we eliminated every hand-engineered feature' (phoenix/README.md). No tweepcred, no RealGraph, no follower-ratio feature enters the 2026 scorer.",
      "code": "grep of xai-org/x-algorithm: zero hits for tweepcred/realgraph. phoenix/recsys_model.py RecsysBatch = ID hashes + action history + post age only."
    },
    {
      "slug": "external-links-penalized",
      "claim": "\"The algorithm nukes posts with external links\"",
      "verdict": "NO RULE IN THE CODE — but independent studies measure a real penalty anyway",
      "explain": "There is no link-detection or link-penalty term anywhere in the open 2026 scoring path — in plain words: nobody wrote a rule that docks links. What's real: links change which heads you earn (click vs dwell), and low predicted click-dwell on clickbait leaks value. Two ways a demotion could still exist unseen: the transformer could LEARN link-correlated demotion from engagement data with no written rule anywhere (learned behavior is invisible in code by design), and trust-and-safety visibility rules live in a withheld crate where a link policy could sit. And the behavioral evidence is now strong: Galeazzi et al. (NDSS 2026, 40M+ tweets) find link posts get 4.67-7.45x lower median visibility, with a gradient by link type (news links worst); Efstratiou et al. (ICWSM 2026, ~205k exposures in paired counterfactual feeds) find external links the strongest negative predictor (p<0.001) while follower count is NOT significant (p=0.14) - so it isn't an account-size artifact - and find VIDEO EXEMPT. Buffer's 18.8M-post sample shows non-Premium link posts falling to 0% median engagement after March 2025. X's leadership denies all of it (Nikita Bier: 'links were never deboosted'; Musk, Jul 29 2026: 'we haven't for over a year'). Honest answer: no written rule in the open ranker, but a large measured effect in the wild - which is precisely the learned-or-withheld case, and the strongest reason to distrust anyone who answers this question with a flat yes or no.",
      "code": "home-mixer/scorers/ranking_scorer.rs (no link term); vf_filter.rs → withheld xai_visibility_filtering crate."
    },
    {
      "slug": "engagement-pods",
      "claim": "\"Engagement pods / early-boost groups beat the algorithm\"",
      "verdict": "MECHANICALLY WEAKENED by the 2026 design",
      "explain": "Phoenix predicts per-viewer engagement probabilities from each viewer's own history — candidates never see each other's engagement counts in the scoring pass (candidate-isolation attention mask), so a pod's likes don't raise other viewers' predicted P(action) directly. Meanwhile Grox runs LLM spam screens including low-follower reply-spam detection. Pod likes are also the cheapest head (2023: 0.5).",
      "code": "phoenix/README.md (candidate isolation); grox/classifiers/content/spam.py; grox/plans/plan_spam_comment.py."
    },
    {
      "slug": "golden-hour",
      "claim": "\"You must get engagement in the first 30/60 minutes or the post dies\"",
      "verdict": "PARTLY REAL, WRONG MECHANISM",
      "explain": "There is no hand-coded freshness decay in the 2026 ranker. Post age enters as a learned feature bucketed in 60-minute steps, capped at 4800 minutes (80 hours) — after which all posts look equally old — and an AgeFilter hard-drops old posts. We then MEASURED the learned curve by running the released checkpoint at 15 impression ages (see the Lab page): predicted replies collapse to ~36% of fresh by day 3, reposts peak at 2-4 hours (~1.5x) then fade, predicted favorites never decay at all, and every curve flatlines at exactly 80 hours — the bucket cap, visible in behavior. So: no first-hour cliff, but a real per-action conversation window.",
      "code": "phoenix/recsys_model.py POST_AGE_MAX_MINUTES = 4800; home-mixer/filters/age_filter.rs."
    },
    {
      "slug": "hashtags",
      "claim": "\"Hashtags boost your reach\"",
      "verdict": "NO HASHTAG FEATURE in the 2026 ranker",
      "explain": "The ranking model's inputs are ID-hash embeddings, engagement history, surface, and post age. Post text (including hashtags) reaches the system only through Grox multimodal embeddings — where a hashtag is just more text, and spammy tag-stuffing is exactly what LLM spam screens catch.",
      "code": "phoenix/recsys_model.py RecsysBatch; grox/embedder/multimodal_post_embedder_v5.py."
    },
    {
      "slug": "post-more",
      "claim": "\"Post 10+ times a day for maximum growth\"",
      "verdict": "DISCOUNTED IN CODE",
      "explain": "The author-diversity decay multiplies your Nth-best post in each viewer's session by (1−floor)×decay^N+floor — your own posts compete with each other, per viewer, every session. Volume also raises predicted mute/not-interested. More posts help until the decay + negative heads outrun them; the curve is in the code, its constants are not.",
      "code": "home-mixer/scorers/ranking_scorer.rs apply_author_diversity()/diversity_multiplier()."
    },
    {
      "slug": "tools-run-the-real-algorithm",
      "claim": "\"Tool X scores your draft with X's real algorithm\"",
      "verdict": "IMPOSSIBLE with what's public — anyone claiming it is guessing",
      "explain": "Two hard blockers, from the code: (1) every 2026 weight lives in an unpublished params service; (2) the released mini-Phoenix checkpoint consumes learned ID-hash embeddings — a brand-new draft has no trained embedding, so 'running the model' on your draft produces noise. We went further and measured the ceiling (see the Lab page): even on posts the checkpoint WAS trained on, its predictions rank an author's own posts against their actual results at a median correlation of 0.00 — a coin flip — because what it actually encodes is per-viewer personalization (a 6.7x swing from three history items), not post quality. What CAN be computed honestly: the exact score structure, the hard gates and filters, and clearly-labeled 2023 weights. That is what this site does, and the measured ceiling of what any tool can do.",
      "code": "home-mixer/scorers/ranking_scorer.rs (crate::params withheld); phoenix/recsys_model.py (ID-hash inputs); phoenix/artifacts (frozen mini checkpoint)."
    }
  ]
}