What the work actually involves

You'll spend most sessions in three modes: writing a reference artifact the way you'd produce it for a real company, prompting the model to produce its own version, and then documenting precisely where the model's output would fail in practice. The artifacts are the ordinary furniture of a people function — a scorecard for a Series B ops hire, a handbook section on remote-work expense reimbursement, a rejection email after a final-round loss, a headcount tracker with comp bands and backfill logic. The judgment Ethos is buying is the part that isn't in the template: knowing that an offer letter that states an annual salary without clarifying at-will status invites trouble, or that a performance-review form with a five-point scale and no behavioral anchors will produce ratings inflation by the second cycle.

Spreadsheets and slides count as much as prose. A comp-band tracker that looks right but hard-codes a midpoint instead of computing it from range min and max is a real failure mode, and you're expected to catch it and say why. Similarly for a board-style headcount slide that buries the ask.

What the platform screens for

  • Direct ownership, not proximity. Screens probe whether you wrote the policy or routed someone else's draft. Expect follow-ups on specific documents you personally authored.
  • Failure articulation. Anyone can say an AI draft is "generic." The screen rewards people who name the consequence: which candidate this loses, which claim creates legal exposure, which clause contradicts the one three pages up.
  • Tool craft under scrutiny. Formula structure, conditional formatting logic, slide hierarchy — asked about concretely.
  • Calibrated availability. Consistent weekly hours beat an ambitious number you won't hit.

Logistics

Fully remote and fully asynchronous; there are no standing meetings and no timezone requirement. Work volume fluctuates with the lab's task queues, so the 5–20 hour band is a range rather than a commitment in either direction. The $70/hour figure is what contributors on this track have reported; Ethos sets rates per project and it is not guaranteed. The AI voice screen runs roughly 20–30 minutes and is conversational — it asks follow-ups based on what you just said, so prepared monologues tend to come apart.