Method
A simple rating from blind choices.
The first version uses standard Elo-style updates. It is meant to be understandable and easy to improve as the vote pool grows.
Same prompt
Two available models receive the same fantasy or RPG prompt through CharGen’s normal generation paths.
Hidden identity
Voters see candidate A and candidate B. Names are not in the ballot response until a non-skip vote is recorded.
Rating update
Models begin at 1000. A win, loss or tie adjusts both ratings with a K-factor of 32.
Counted decisions
What changes a rating
A wins, B wins and ties affect Elo. “Both poor” is retained as product feedback but does not award either model a win. Skips do not affect ratings.
Ratings are preference signals, not objective quality scores. Small samples can move quickly. Use the vote count beside each model when reading the table.