The work

A frontier AI lab is building evaluations of how well models read real radiology studies, and Mercor is recruiting practicing radiologists to define the reference standard. Day to day this means opening a de-identified study, writing the reference report or isolating the specific finding under test, and applying the structured system the case calls for — RECIST target selection and measurement, a Lung-RADS or BI-RADS category on a borderline lesion, CAD-RADS on a coronary CTA, ASPECTS on an early infarct. A second stream of work is adjudication: you read an AI-drafted report alongside the images and mark what is wrong, what is asserted without imaging support, and what would change management if a clinician acted on it. Contributors are also asked to bring cases from their own reading that a strong model would plausibly get wrong.

What the screen is looking for

This is not annotation work and the screening reflects that. Expect to be pressed on measurement convention — axial versus true short axis, where calipers go on a spiculated mass, which nodes qualify as measurable — and on interval comparison against a prior of different technique or slice thickness. Expect scenarios where the model's report is fluent and internally consistent but wrong, and you are asked what you would flag and how you would justify it in writing. Vague answers collapse quickly under follow-up; the screener is calibrating whether you reason from the pixels and the published criteria rather than from report boilerplate.

Logistics

  • Fully remote and asynchronous; work is pulled from a queue rather than scheduled in shifts.
  • Volume is flexible — many contributors commit 5–10 hours a week alongside clinical practice.
  • Observed rates for this listing run $300–400/hr, varying by sub-specialty scarcity and task type. Rates are as reported, not guaranteed.
  • Board-certified or board-eligible radiologists (ABR, FRCR, EDiR, or international equivalent); later-stage residents and fellows are sometimes considered. Current or recent active reading is expected.