What the work actually is

You write compliance problems hard enough that a capable language model gets them wrong — and then you explain, in writing, exactly why it was wrong. A task might be a suspicious-activity pattern that looks like structuring but isn't, a conflicts-of-interest disclosure that satisfies the letter of a policy while defeating its purpose, a marketing review under FINRA 2210, a HIPAA breach-notification timing question, or a third-party due diligence file with a gap a regtech screen wouldn't flag. Each item needs a defensible reference answer, the regulatory citation or supervisory guidance behind it, and a rationale a reviewer who isn't a compliance officer can follow.

The second half of the job is evaluation: reading model output and scoring it. The interesting failures are rarely factual — models are good at reciting rule numbers. They fail on judgment: escalating what should have been documented and closed, treating guidance as binding, giving categorical advice where the answer depends on the firm's risk appetite, or writing a policy that no first line could actually operate. Your feedback has to name the failure mode, not just mark it wrong.

What the screen looks for

  • A named regime you've actually operated under. BSA/AML, FCPA, HIPAA, GDPR, SEC/FINRA, CMS, OSHA, EU MDR — depth in one or two beats a tour of all of them.
  • First-hand process experience: risk assessments you built, testing or monitoring you ran, investigations you closed, exams or audits you sat through. Second-hand familiarity shows immediately under follow-up.
  • The ability to be wrong precisely — to say where a rule is ambiguous, where reasonable compliance officers disagree, and what you'd need from the business to decide.
  • Written clarity. Nearly all output is prose; the screen is text-heavy for that reason.

Logistics

Fully remote and asynchronous, contractor engagement, no fixed hours. Most contributors work in blocks — a few hours of authoring, then a batch of evaluations — and throughput matters more than schedule. Expect a calibration round before volume work: your first items get reviewed against a rubric and you revise. Pay is hourly in an observed $50–100 band; the top of it tends to go to people with a certification (CRCM, CCEP, CAMS) or a JD plus real supervisory or examination experience. Nothing about the rate is guaranteed, and volume varies by project.