January 2026: xAI published the full X recommendation algorithm — pipeline code (Rust/Python), the Grok-architecture Phoenix ranker, ads blending, content classifiers — with a four-week public update cadence, and by May a runnable end-to-end pipeline plus a downloadable mini-Phoenix checkpoint. Withheld: the production weights and the live-tuned constants; and there is no way to verify the GitHub tree matches what the servers run.
The legibility axis
industrialization-needs-a-mechanical-arbiter said industrialization requires a mechanical judge. This event exposes the finer structure: judges differ in legibility, and the ease of industrializing against them is monotone in it.
Each step right cheapens the extraction layer: MrBeast needed a decade of dissection, Jenny a crawler and readability tests, the GEO (generative-engine-optimization) practitioner an interrogation loop — an X grower now needs git clone.
What dies and what survives
The constitution went public; the judiciary stayed private. Rules — stage structure (candidate sourcing → Phoenix ranking → filtering), which signals enter as features, hard gates, diversity caps — are now free. So rule-arbitrage dies: the sellable meta-knowledge of "how the judge works" (the coaching industry's X chapter, growth folklore) is expropriated to zero. Taste-arbitrage survives: what Phoenix will score this artifact remains a black box, because weights and live constants stayed home. The residual list from the-mrbeast-factory §6 (position, capital) gains a member: whatever the code cannot say.
Red-Queen corollary: when every producer reads the same rulebook, rule-knowledge stops being differential. The floor rises (anyone can SOP-ize), and the relative value of the unpublishable — idea supply, taste, capital — rises with it.
Reading the code (repo cloned 2026-07-16, ~/Develop/reference/x-algorithm)
The withholding is more surgical than the press coverage says. home-mixer — the orchestration layer — does not compile as published: a dozen files use crate::params, but no params module exists anywhere in the repo and lib.rs never declares one. The excised file is exactly the price list: the 19 engagement weights, AUTHOR_DIVERSITY_DECAY/FLOOR, MIN_VIDEO_DURATION_MS, the negative-score offsets, result sizes, feature flags. They did not even ship placeholders — a placeholder would leak magnitudes.
What the form nonetheless reveals (all in home-mixer/scorers/):
- The action vocabulary — final score = Σ weightᵢ·P(actionᵢ) over: favorite, reply, retweet, quote, click, quoted-click, photo-expand, profile-click, video-quality-view (gated: only counts above a minimum video duration — threshold withheld), share, share-via-DM, share-via-copy-link, dwell, continuous dwell time, follow-author, and four negatives: not-interested, block-author, mute-author, report. Dark-social sharing is a first-class judged action; so is "made the viewer check the author".
- Author diversity — repeated authors in one response decay exponentially to a floor:
(1−floor)·decay^position + floor. Shape public, strength private. - The hard filters (the only real "rules" left): dedup, age, self-posts, muted keywords, blocked/muted authors, already-seen/served, paywall eligibility; post-selection trust-and-safety and conversation dedup.
- The README's own confession: "We have eliminated every single hand-engineered feature and most heuristics." Relevance is learned end-to-end from engagement sequences — so beyond the filters and the score assembly, there were never many rules to publish. The judge is almost all taste; the mini-Phoenix checkpoint is a frozen snapshot of a continuously-trained production model.
So the earlier framing sharpens: the constitution shipped with the sentencing table torn out — the disclosure is truncated at precisely the first line that would constrain the judge numerically. What a producer can still take: the sign structure (optimize toward dwell, DM-shares, follow-author; avoid triggering mute/not-interested — the negatives subtract), the existence of the video-duration gate, the filter list (don't die before scoring), and the diversity decay (one author can't stack a viewer's feed).
Why the judge's owner did it
industrialization-needs-a-mechanical-arbiter corollary 3: whoever owns the judge owns the industry's objective function. Why give the rubric away? Context: a transparency fine and platform controversies. The trade: spend the rule-arbitrage (a near-worthless asset to the platform — it mostly fed sophisticated spam-optimizers) to buy legitimacy — an audit surface without auditability (weights withheld; no production-parity proof). The open-sourcing is itself an awareness artifact: the judge became content, on a four-week release schedule.
PRINCIPLES
- Legibility is the axis industrialization slides on: opaque → inferrable → interrogable → published; each step collapses more of the craft into optimization.
- Publishing the form is not publishing the taste: architecture open + weights closed = rules expropriated, judgment retained. Read what was withheld to see where the power stayed.
- Open rulebooks equalize downward: rule-knowledge stops being a moat the moment everyone holds it; differential advantage migrates to what cannot be published.
- Transparency can be packaging: an unauditable disclosure is an awareness play by the judge itself — grade it on verifiability, not on the gesture.
Sources: xai-org/x-algorithm · May 2026 release: runnable pipeline + mini-Phoenix, production weights withheld · what the code says / verifiability gap · TechCrunch: released amid transparency fine · pipeline mechanics