Developer improves 20k-star ML repo by measuring a bias, not fixing it

A developer building a Japanese decision model tested the open-source laya multilingual checkpoint against 300 label-conditioned Japanese business emails. The evaluation revealed that laya's ordinal scoring head almost never selected the first-listed option, regardless of wording or order, pointing to a positional bias in the model weights. Rather than attempting a code fix, the contributor opened an issue with reproducible measurements and an A/B test that a third party helped trace to the checkpoint weights. Within four days, the maintainer documented the limitation in a new release and merged a regression-check tool the contributor had written. The underlying fix requires retraining the model, which has not yet occurred, but the contributor's benchmarking work established a clear, verifiable baseline for evaluating future improvements.