A contender isn't a model — it's a pipeline. A deslop prompt running on a particular base model, a third-party humanizer behind an HTTP endpoint, or nothing at all as a control. Prompt and model stay separate, so the leaderboard can answer whether a given prompt wins everywhere or only on one model.
Elo is path-dependent: it updates incrementally, so the same votes in a different order give different ratings, and the result depends on a learning rate nobody can justify. Bradley–Terry is a maximum-likelihood fit over the whole vote set at once. The same votes always produce the same numbers.
Each score carries a 95% bootstrap interval. When two systems' intervals overlap, the votes collected so far don't reliably separate them — the ordering between those two is noise, not signal. Treat the bars as the real result and the single number as a summary.
A tie counts as half a win to each side, the standard Bradley–Terry treatment. “Both are bad” is stored but excluded from the fit: it says the input defeated both systems, which is information about the text rather than a preference between the pair.
Preference for deslopped text is badly confounded by brevity — cutting a third of the content reads as an improvement whether or not it was one. Every rewrite's character and word count is logged from the first vote, because that can't be reconstructed later.
System names are hidden until you vote, and which rewrite appears as A is randomised per battle, so position bias stays measurable. Voting is anonymous with no account, rate-limited per session, and one vote per battle.