Two athletes. One choice.

1. The arena serves you a pair — usually two athletes from the same sport with similar ratings, because that is the comparison that actually tells us something.
2. You choose. Click a panel, press ← or →, or swipe.↓ skips a pair you have no opinion about, which is a useful answer too.
3. Both ratings update immediately, and only then are you shown how everyone else voted on that same pairing. The split is deliberately withheld until after your choice — showing it first would measure the anchor rather than your opinion.

No account is needed and there never will be one on the way to voting. Signing in only adds things: your own ballot, and the ability to follow athletes.

Every athlete carries a Glicko-2 rating and, just as importantly, a rating deviation — how sure the system is of that rating. Everyone starts at 1500 with a deviation of 350, which is to say: a complete unknown.

That second number is why beating a well-established athlete moves you further than beating someone nobody has voted on, and why an upset moves the board more than a predictable result. It is also why we will not print a rank next to a rating we do not trust.

An athlete is provisional until they have 30 matchups and a deviation at or below 110. Provisional athletes are listed separately and carry no rank number at all — not a low rank, no rank. The full reasoning, the two rating tracks and the exact configuration are on the methodology page.

Cross-sport votes move an athlete’s global rating — a fan-preference index across all 27 sports. Only same-sport votes move their sport rating, which is what the per-sport boards show. A boxer beating a golfer says something about reputation; it says nothing about boxing, so it is not allowed to move a boxing ranking.

Every vote is authorised by a signed, single-use token tied to the exact pairing you were shown, so a vote cannot be submitted by anything that never loaded the page, and cannot be submitted twice. Rate limits cap how fast one person can go.

Voting that looks automated — machine-speed clicking, a fixed side pattern, one athlete taking an implausible share of a session — is quarantined: the vote is kept in the record, excluded from the maths, and reviewed. Nothing is ever deleted, and because every rating is recomputed from the vote log rather than patched, a review that clears a batch restores the ratings exactly.

What that means in practice: no single vote can move a board on its own, and a coordinated push can be undone without collateral damage to everyone else’s votes.

It is not a performance model. Nobody here is simulating a fight or adjusting for era. The board measures which athletes a crowd of fans prefers, which is a real and interesting thing, and a different thing from who was better.