Benchmark builder
Create a medical AI benchmark through seven guided planning stages. Define the clinical question, choose the data, and design the scoring rubric, then export a reviewed plan and instructions to build and run it locally.
Build a case study
Author a single benchmark item: the clinical case, the expected answer, and the rubric a judge would score against. The unit of evaluation that benchmarks are made of.
Annotate benchmark items
Label ground-truth answers and grade model outputs against a rubric. Multi-rater concordance, disagreement adjudication, and LLM-suggested consensus all happen in one surface; exports annotations.md.
Dormant identifier-signal hook
Inputs are assumed stripped. An optional background heuristic records only advisory states and coarse categories. It does not redact, block, certify, or clear data.