What the work actually looks like

You will spend most of your time in two modes. In the first, you author prompts against a provided dataset — a CSV, a spreadsheet, a small relational extract — designed to make a model retrieve, join, filter, or reason over that data and fail in an informative way. In the second, you read a model's response and decide whether it is factually correct, logically sound, and complete, then write the justification: which cell, row, or clause in the source proves the point. Hallucinations, silent arithmetic errors, dropped conditions, and answers that are right by luck but wrong by reasoning are all separate failure categories, and the annotation guidelines will expect you to distinguish them rather than collapse everything into "bad response."

What the screen is looking for

Turing gates on a master's degree or higher in any discipline plus at least three years of professional, research, or teaching experience — the degree requirement is stated flatly, so expect it to be verified. Beyond that, the assessment is about whether your judgment survives contact with ambiguity: can you tell a reasoning gap from a formatting complaint, do you cite evidence rather than assert, and do you apply a rubric the same way on your fortieth item as on your fourth. Prior exposure to LLMs, prompt engineering, or annotation work is preferred rather than required, but candidates who can describe a specific model failure they found and how they proved it tend to clear the bar faster than those who speak about AI in general terms.

Logistics

  • Contractor assignment, four-week duration, no medical or paid leave
  • 20, 30, or 40 hours per week — you choose a band and are expected to hold it
  • Remote and largely asynchronous, but at least four hours per day must overlap Pacific time
  • Pay is undisclosed in the listing; Turing states "competitive compensation" without a band, so treat any figure you hear secondhand as unverified
  • Shortlisting is based on qualifications plus assessment scores, so the written assessment carries real weight