Physician baseline

Rubric scores carry no natural interpretation on their own. A score of 0.5 means half the available weighted criteria were met, which says nothing about adequacy until a human reference is scored the same way. The requirement is that the baseline responses pass through the identical grader and adjustments; a baseline collected under a different procedure is not a baseline for that scale.

HealthBench reports a graded progression on the parent set. Physicians writing unaided scored 0.13. Physicians given September 2024 model responses as references scored 0.31. Physicians given April 2025 model references scored 0.48, at which point they were no longer improving on the references, which scored 0.49 on their own. On HealthBench Professional, physician-written responses score 0.437 against the same rubrics, graded by GPT-5.4 at low reasoning effort with a length adjustment.

These numbers are frequently misread. A rubric awards credit for stating material that a clinician writing to a colleague would leave implicit, so an unaided score of 0.13 measures rubric coverage under the grading procedure rather than clinical competence. The reference-assisted sequence is the more informative reading, because it marks the point where model output stopped adding to expert output on that set. See grader bias for why coverage and quality diverge.

A model scoring above a physician baseline has exceeded a specific human reference on a specific rubric under a specific grader. That is not a deployment clearance, and it is not evidence about the tasks the rubric does not cover. Current per-model figures for the HealthBench family are on healthbenchprofessional.com and clinicalbenchmarks.ai.