Auditing eleven unanswerable task variants
Six initial variants needed correction before their refusal results could be interpreted.
Experiment notes
Research notebook
Questions, experiment setups, results, failed approaches, and next steps.
Design correction
Task validity, calibration, and judge disagreement
Historical BiomniBench-DA experiments on contamination, cohort design, workflow consistency, refusal, and scoring.
Elo and correctness on GPQA-bio
Sample size, aggregation, and uncertainty in the historical Elo–accuracy analysis.