What the work actually looks like

Most days you will be doing one of three things. First, authoring problems: writing a question in your field that is unambiguous, has a verifiable answer, and is hard enough that a strong model gets it wrong — then documenting the full worked solution so a reviewer can check your reasoning line by line. Second, grading model output: reading a model's multi-step derivation or experimental reasoning and deciding not just whether the final answer is right, but where exactly the chain broke. Third, adjudicating disagreements between other experts' labels, which means defending a position in writing against someone with comparable credentials.

Tasks arrive in batches tied to specific projects — a chemistry batch on reaction mechanisms, a physics batch on statistical mechanics, a math batch on real analysis proofs. Batches open and close, so throughput is uneven; expect stretches of steady work followed by gaps while a new project spins up. Rubrics are usually detailed and occasionally ambiguous, and part of the job is flagging when a rubric doesn't cover the case in front of you rather than guessing silently.

What the screen looks for

  • Verifiable depth in a named subfield. "Physics" is not a credential; spin glasses, cavity QED, or Landau theory is. The screen pushes on specifics until it finds the edge of what you actually know.
  • Whether you can be wrong on purpose. Good problem authors know which mistakes a plausible-sounding model makes, and can construct distractors that punish pattern-matching rather than reasoning.
  • Calibration. Saying "I'd escalate, I can't verify this claim" scores better than confidently defending a shaky answer.
  • LaTeX and written clarity. Solutions are read by other experts; sloppy notation costs you.

Logistics

Fully remote and fully asynchronous — no standing meetings, no timezone requirement. Mercor typically asks for a minimum weekly commitment (often 10–20 hours) and reliability matters more than volume; missed batch deadlines end engagements faster than slow output does. Pay is hourly at the observed $75/hr band for this listing, though rates vary by domain, seniority, and project, and are not guaranteed. Identity verification and a live or recorded technical screen are standard.