Methodology

TrialDesignBench is built around clinically meaningful, reproducible evaluations.

Task Construction

Tasks should be grounded in realistic trial design scenarios and reviewed by domain experts in clinical research, biostatistics, or regulatory science.

Evaluation

Each task should be decomposed into checkpoints that capture the critical parts of the workflow: correct interpretation of evidence, appropriate endpoint selection, defensible statistical assumptions, safety-aware eligibility criteria, and coherent documentation.

Reproducibility

Evaluation runs should use versioned task inputs, pinned harness dependencies, and recorded trajectories so results can be audited and compared over time.

Safety

High-stakes clinical trial design requires conservative grading. Unsafe recommendations, unsupported assumptions, missing safeguards, or incoherent protocol elements should prevent a task from passing even when other checkpoints are correct.