Home › Methodology

Methodology: what’s knowable, what isn’t, and how we label both

This site exists because the X-growth industry answers a well-defined technical question — “how does the ranking pipeline treat my post?” — with folklore. Here is our entire epistemology, in public.

The three tags

What nobody outside X can do (including us)

  1. Quote current weights. All 22 constants load from crate::params at runtime — the module is not in the repo (home-mixer/scorers/ranking_scorer.rs).
  2. Score your draft with the real model. The released mini-Phoenix checkpoint consumes learned ID-hash embeddings (phoenix/recsys_model.py) — a new draft has no trained embedding, so “running the model on your post” produces noise. The Rust services also ship without Cargo.toml or several private crates: reference code, not a buildable system.
  3. Read your trust-and-safety status. Visibility enforcement lives in the withheld xai_visibility_filtering crate (home-mixer/filters/vf_filter.rs). Shadowban-checker tools are inferring from behavior, not reading state.
  4. See Grox’s prompts. The LLM classifiers’ prompt templates and configs are withheld imports — we know what is screened (spam, safety, “banger” quality), not the criteria.

What we therefore do

We compute only what the code makes computable: the exact scoring structure (22 terms, negative offset, diversity decay, OON multiplier, video-duration gate), the filter pipeline, and clearly-dated 2023 numbers for direction. Where we add judgment — slop phrasing, bait patterns, commonly-muted terms — the tag says so. When a claim can’t carry a tag, we cut it.

Why “the code is public” is not the same as “the code is complete”

There is one documented case, and it is worth knowing before you trust any code dump — X’s or ours. In February 2023 Platformer (Zoë Schiffer and Casey Newton) reported that engineers had built a system boosting Elon Musk’s own posts by an artificial factor internally called the “power user multiplier.” Six weeks later the 2023 open-sourcing shipped with a feature literally named author_is_elon — spotted by Jane Manchun Wong, reported by TechCrunch on March 31 2023, and removed from the repository within hours of discovery. Musk’s own comment that day was that the code was “going to be quite embarrassing.”

That is the whole argument for treating a published repo as evidence rather than proof: the release corroborated the reporting, and the response was deletion, not explanation. It is also why this site tags claims by where they come from instead of asserting that the open code is the system. The 2026 release withholds more than the 2023 one did (every weight), so the gap between “published” and “complete” is wider now, not narrower.

Standing corrections policy

The repo updates (~4-week cadence, per xAI); we re-verify each drop and log it on the changelog. If a code change invalidates a page, the page changes the same day and says it did. Found an error? The fastest correction channel is a public post — we read the quote-posts.