Existing work, made inspectable

Twenty-five questions. The actual prompts, answers, and judge scores.

This page shows what the finished exploratory attempt produced—without rerunning models or smoothing over the failed row.