What the work actually is

This is authoring, not rating. For each task you produce three things: a simulation the model can query (Python — typically SciPy/NumPy, SPICE via a wrapper, or a lumped-parameter model you write yourself), a specification sheet where every requirement is a numeric threshold with a clear pass/fail rule, and a prompt that states the design problem without leaking the answer. The model gets a bounded number of calls into your simulation, probes it, and submits a final design — gains, component values, geometry parameters. An agentic grader executes your sim against that submission and scores spec satisfaction, so your thresholds have to be unambiguous enough to survive automated evaluation without a human arbitrating edge cases.

The interesting difficulty is calibration. A task the model solves by pattern-matching a textbook second-order compensator teaches the lab nothing; a task nobody could solve because the spec set is internally contradictory is a defective item. You are looking for problems where the path runs through actual reasoning — coupled trade-offs, non-obvious stability or thermal limits, specs that conflict until you find the right operating point — and where you can defend correctness because you designed the physics, even if you deliberately built something you couldn't shortcut by inspection.

What the screen looks for

Mercor's screening is AI-led and follow-up heavy. Expect it to pick one domain claim from your background — a loop you compensated, a PA you matched, a converter you stabilized — and push two or three layers into the numbers: bandwidth, phase margin, ripple, ambient assumptions, why that topology. Vague seniority does not survive this. It will also probe whether you can convert engineering judgment into machine-checkable criteria, and whether you understand why a sim that is trivially gameable (unbounded actuator authority, no noise, no component tolerance) is a bad task.

Logistics

  • Fully remote and asynchronous; you pick your hours and task volume, with onboarding calls offered daily while you ramp.
  • Work happens in a browser-based studio plus GitHub for simulation code — expect to commit, review, and iterate on your own tasks.
  • Observed band is $60–90/hr, set by domain and depth of track record; rates are as reported by contributors, not guaranteed.
  • Prior model-evaluation experience is not expected — the scarce credential here is the engineering depth and the Python fluency to encode it.