Hamed Khosravi
← All publications

Certified Corruption Budgets: Anytime-Valid Leaderboard Claims under Adaptive Rigging

Hamed Khosravi, Xiaoming Huo

arXiv preprint arXiv:2610.10597, 2026

Certified corruption tolerance on real Chatbot Arena data: for each ordered pair, how many corrupted records the claim
Certified corruption tolerance on real Chatbot Arena data: for each ordered pair, how many corrupted records the claim “the row model has had an edge over the column model” can withstand — simultaneously over all 16,512 ordered pairs of 129 models. Flipping votes already cast costs an attacker roughly half as many records as forging new ones. Empty cells are not certified; dots mark pairs with no battles.

The problem

Because every standing on a public leaderboard is visible, an adversary can watch each claim as it is published and rig the votes behind it — forging ballots, withholding private variants, or contaminating the benchmark. A guarantee that assumes the data were collected before anyone looked does not survive that.

What we do

Each claim is published together with a certified corruption budget: a tolerance such that the claim is correct unless more records than that were corrupted. The certificate holds against attackers who see every certificate and are not themselves held to a budget.

What it buys you

A reader gets a number attached to each claim saying how much tampering it would take to overturn it — which is a far more useful thing to know about a public leaderboard than a p-value computed as if nobody were watching.

Cite

@article{khosravi2026certified,
  title   = {Certified Corruption Budgets: Anytime-Valid Leaderboard Claims under Adaptive Rigging},
  author  = {Khosravi, Hamed and Huo, Xiaoming},
  journal = {arXiv preprint arXiv:2610.10597},
  year    = {2026}
}