What the work involves

You receive Kubernetes tasks — cluster-operations scenarios, manifest sets, incident-troubleshooting prompts — that are being used to train or evaluate frontier models, and you decide whether each one is correct and worth keeping. That means reading the manifest for errors the author didn't intend, checking that the stated failure mode is the one the setup actually produces, confirming the reference solution is the right fix rather than a coincidental one, and flagging tasks where multiple valid answers exist but the rubric only accepts one. A CrashLoopBackOff scenario that's really an image-pull problem, an RBAC task whose role binding grants more than the prompt implies, a PVC exercise that silently depends on a specific storage class — these are the defects you're paid to catch.

Each audit ends in written feedback against a rubric: a score, the specific line or resource that caused it, and what would have to change for the task to pass. Volume matters, but so does consistency — your scores are compared against other auditors' on overlapping items, and unexplained divergence gets reviewed.

What the platform screens for

  • Depth under follow-up. Mercor's screen is AI-led and will keep asking. Naming `kubectl describe` is not an answer; explaining what in the events and container statuses distinguishes an OOMKill from a liveness-probe restart is.
  • Real production exposure. Expect questions anchored to incidents you personally handled — cluster, scale, what broke, what you changed.
  • Evaluation judgment. Whether you can separate "this task is wrong" from "this task is written in a style I wouldn't use," and whether you can justify a score in writing that another reviewer would reproduce.
  • Code literacy. Enough Go, Python, or TypeScript to read operator code, controllers, and test harnesses attached to tasks.

Logistics

Remote and asynchronous, contractor engagement, work pulled from a queue rather than assigned in shifts. Pay in the $70–90/hr band has been observed for this role; rates vary by assessed depth and are not guaranteed. Most contributors commit 10–20 hours a week, and sustained throughput usually matters more than availability at particular hours. Some projects require a machine capable of spinning up a local cluster (kind, minikube, or a sandbox cloud cluster) to reproduce scenarios.