How the Proving Ground works
Ratings come from blind, side-by-side votes on the same prompt. Here is exactly how a vote becomes a score.
1.Blind by design
You write one prompt. Two models are drawn at random from those available on your plan and each makes one result. Their names are hidden and their left and right positions are random, so the only thing to judge is the work.
2.One question per vote
Every vote answers a single, named question — for now, “Which one matches the prompt better?”. You can pick A, pick B, call a tie, or say both missed. Each battle takes one vote, and the models are revealed only after it is counted.
3.From votes to a score
Scores use an Elo rating. Every model starts at the same score. After a vote, the winner gains points and the other side loses the same amount; an upset against a higher-rated model moves the scores more than an expected win. A tie counts as half a win for each side. “Both are bad” changes no scores but still counts as a vote.
4.What is not scored
Price, speed and maximum resolution are left out. A cheap, fast model and an expensive, slow one face the same question. Only active models with at least one vote appear on a board.
5.What a battle costs
Each side is a normal generation charged from your credits, at half price when the Proving Ground discount is on. The range is shown before you start, and the exact amount is held when the battle begins. If a side fails, its credits are returned.
The formula
Expected score E = 1 / (1 + 10^((Rb − Ra) / 400)). New rating = Ra + K × (S − E), with K = 24 and S = 1 for a win, ½ for a tie and 0 for a loss.
E = 1 / (1 + 10^((Rb − Ra) / 400))Ra′ = Ra + 24 × (S − E)Worked example
A model rated 1,500 beats one rated 1,600. It was expected to score 0.36, so it gains 15 points and moves to 1515. The other model loses the same 15 points.