AI Leaderboard Governance: A 56-Board Audit of Submission, Retesting, and Score Removal Rules

Background. A leaderboard gives a score, but not necessarily the rules behind it: who could submit a model, whether variants were tested privately, how uncertainty was reported, or when a score could be removed or corrected. Methods. I built a source-traced registry of 56 public AI leaderboards from the operators' own pages (source rows checked 2026-10-20, probes 2026-10-24). I coded five procedural disclosure dimensions and three further cells as stated or not found in the material searched. Results. A submission route or rule was stated on 32 boards and a contamination mechanism on 24. A repeat requirement was stated on 7 and uncertainty reporting on 20 (21 in all). Removal or correction rules appeared on 6 (4 prospective removal rules, 1 correction rule, 1 discretionary disqualification). Some private-testing, exposure, publication or submission-frequency disclosure appeared on 7, including 4 that describe private pre-release variants. In the searched material, 9 boards stated none of the five dimensions. Validation. I checked quotations verbatim against saved pages; every one of the 90 stated cells across the five dimensions has a cell-specific passage. The classifications remain provisional. AI quote-window agreement, measured against pre-adjudication labels, varied by dimension. An independent AI adjudicator reading excerpts agreed with 43 of 50 sampled labels (5 undecidable; the two disagreements were recoded). Human and whole-page validation are not complete. Contribution. The Leaderboard Governance Registry makes these disclosures, their sources and the search boundaries available as a versioned, machine-readable record. It does not measure policy enforcement or ranking integrity.