What the work actually is

This is authoring, not annotation. You choose your own source material — a paper you know well, an open dataset, an open-source solver, or a scenario you design from scratch — and turn it into an executable research problem with a defensible ground truth. Each task needs a prompt that is unambiguous to a competent mathematician, a reference solution that runs, and grading criteria precise enough that a model's output can be scored without argument. You then calibrate: run the task against frontier models, and it only ships if strong models fail it more often than they pass. Tasks that models solve on the first try get reworked or discarded, so expect iteration to be part of the normal rhythm rather than a sign something went wrong.

The mathematical range is narrow and deep. Mercor is looking for demonstrated depth in at least two of numerical linear algebra, computational mechanics, and computational finance — the kind of depth where you know which conditioning pathologies break a naive implementation, where a preconditioner earns its cost, or why a Monte Carlo variance-reduction scheme fails on a particular payoff. Problems that are merely hard to compute are less valuable than problems where the reasoning about method choice, stability, or convergence is where models break down.

What the screen looks for

The process starts with a resume and application form, followed by a 25-minute conversational interview on background, experience, and motivations. Expect the interviewer to push past your thesis title into what you personally implemented and why. Claiming two subdomains means being ready to talk about both concretely — the specific solvers you wrote, the failure modes you debugged, the papers whose numerics you reproduced. Vague coverage of three areas screens worse than exact command of two. Tooling questions are real filters here too: authoring runs through pull requests with automated quality checks, so Git branching, code review, and getting a reproducible environment running in Docker are working requirements, not nice-to-haves.

Logistics

  • Fully remote and largely asynchronous, coordinated through GitHub
  • Six-week engagement, part-time, 20+ hours per week, immediate start
  • Observed rate for this listing: $70/hr — as posted, not guaranteed, and final rates commonly depend on assessed depth and calibration quality
  • Python for scientific computing is the working language; expect NumPy/SciPy fluency to be assumed
  • Follow-up on next steps typically arrives within a few days of the interview