Eval results

Each cell is the mean over the split with its 95% bootstrap interval (bar). Answer quality is scored by a calibrated Claude judge (rubric v2); retrieval metrics need no LLM. Dev is where settings were chosen; the test split is held out and scored once per milestone.