The effort halo: How LLM judges reward coding style over correctness

LLM judges give higher scores to certain coding styles, independent of whether the code works. We measured the bias, tested causes, and calibrated for it. See how we did it.