About

Instead of asking "which model wins?", I audited whether the human preference signal used to rank LLMs is trustworthy in the first place. I tested whether raters are swayed by surface features — response length, formatting, position, refusal language — rather than answer quality, and quantified how much of the verdict those features alone can explain. I also stress-tested the Zerve agent's own analysis of the same data and documented where it went wrong.