What the work actually is

You pick your own source material — a paper you know well, an open dataset, an open-source repository, or a scenario you design from scratch — and turn it into a self-contained computational physics problem that runs. That means writing the prompt, writing a reference solution that executes cleanly in Docker, and building grading criteria precise enough that another physicist could apply them without asking you what you meant. Then you run the task against frontier models. If they solve it reliably, it does not ship, and you either sharpen the problem or start over. Expect a meaningful fraction of your hours to go into that calibration loop rather than into first drafts.

Authoring runs through Git: branch, commit, open a pull request, respond to automated quality checks and reviewer comments. If your normal workflow is a local notebook you never version, the mechanics will slow you down more than the physics will.

What the screen is looking for

The 25-minute conversational interview probes for genuine depth in at least two named subdomains — condensed matter, optics, quantum information and computing, computational physics, astrophysics, particle physics — and for the coding side of that depth. Being able to describe a Bogoliubov transformation is not the same as being able to say what you would numerically implement, what breaks at the boundaries, and how you would check a result you did not already know. Expect follow-ups that push past your headline answer into method, failure modes, and verification.

  • Bring specific projects, papers, and repositories you can talk about concretely
  • Be ready to name the libraries and simulation methods you actually used, not the ones you've read about
  • Have a clear view of where current models are strong in your area, since that judgment is the job

Logistics

Remote and largely asynchronous. Six-week engagement, part-time, 20+ hours per week, immediate start. The observed rate for this listing is $70/hr; Mercor rates vary by track, assessed depth, and project, and nothing here is a guarantee. Payment follows accepted work, so throughput on tasks that survive calibration matters more than hours logged.