What the work actually involves

You write single-turn prompts in your own practice area and label each one as benign, dual-use or adversarial. A benign prompt might be a routine occupational dose or decay-correction question that a model has no business refusing. An adversarial one sits on the other side of the line — phrased like professional practice, but the underlying intent is source concealment, defeat of detection, or aggregation of material. Then you read what the model returned, score it against a written policy standard, and produce the reference answer: what a correct response looks like, which parts of the question were legitimately answerable, and the technical reasoning for where the boundary falls.

The hard part is not the physics. It is that over-refusal is graded as a failure here, the same as over-compliance. A model that will not discuss attenuation coefficients or Tc-99m elution is useless to the hospital physicist who actually needs it. Your value is calibration — knowing, as the person who held the licence, which questions arrive in an RSO's inbox every week and which ones would have made you pick up the phone.

What the platform screens for

  • Named licence accountability, not adjacency. Mercor is explicit that they want the person of record for acquisition, inventory, transfer and disposal — broad-scope, medical, industrial or irradiator programmes all qualify.
  • Regulatory fluency with NRC, an Agreement State, or an equivalent national regulator: amendments, inspections, findings, incident reporting.
  • Writing under scrutiny. Rationales are read by non-specialists. Prior technical writing, peer-reviewed work, regulatory submissions or expert witness reports are a strong signal; bring a sample or link.
  • Comfort with misuse material. You will read and write about diversion and concealment scenarios for sustained periods. The platform briefs experts on this beforehand and allows pausing without penalty.

Logistics and constraints

Remote and asynchronous, paid per task rather than hourly, in an observed band of $65–75. Task volume on Mercor panels is typically batched — expect irregular availability rather than a steady weekly load, and expect a paid or unpaid calibration exercise before live work. One hard constraint stated up front: nothing you write may involve or draw on classified or export-controlled information, NDA-covered material, or anything subject to prepublication review. Holding such obligations does not disqualify you, but you must disclose them in the application so the work can be scoped around them.