What the work involves

You build the test cases frontier labs use to find out where a model's mortgage advice breaks down. A typical task starts with a borrower fact pattern you author yourself — a self-employed applicant with two years of declining Schedule C income, a VA borrower with partial entitlement and a prior foreclosure, a purchase where the appraisal comes in $18k short a week before closing. You write the reference answer: the correct decision, the agency or program guideline that drives it, and the reasoning chain a competent LO would follow. Then you grade model responses against it.

The grading half is where the judgment shows. Models are fluent about mortgages and frequently wrong in ways that read fine — quoting a DTI ceiling without the compensating-factor context, treating an FHA overlay as agency guideline, giving advice that would put an originator on the wrong side of RESPA, TRID timing, or Reg Z. Flagging plausible-but-noncompliant output is a named responsibility, and the written justification for the flag matters as much as the flag.

What the screen looks for

  • Verifiable licensure. Active NMLS license or registration as an MLO, with an ID that can be checked against the registry.
  • Real origination depth. Expect follow-ups that go past definitions: how you'd document a specific income type, when a program overlay applies, what a disclosure deadline actually triggers.
  • Written reasoning. Reference answers are read by non-mortgage annotators. Vague or jargon-dense writing fails even when the underlying call is right.
  • Program breadth. Conventional plus government and jumbo experience is preferred, not required, but it routes you to more task types.

Logistics

Fully remote, fully async — no calls, no fixed shift. You claim tasks from a queue and set your own hours, with a floor of 10 hours per week. Pay is observed at $70–105/hr, weekly via Stripe; the band typically tracks task complexity and calibration performance rather than tenure. Project work is ongoing rather than a fixed engagement, and volume varies week to week.