What the work involves

You are not writing papers — you are producing the reference material an AI agent is measured against. A task might ask you to resolve a citation collision across a merged BibTeX file, reformat a multi-line equation to a journal's column width, fix a `float` placement cascade, or migrate a manuscript from one document class to another. You execute it in Overleaf or a local toolchain while your trajectory is recorded: the syntax edits, the compile runs, the error messages you read and what you did about them. Clean means deliberate — no flailing, no ten-minute detours you then abandon, and no silent fixes you cannot explain in the write-up.

A second, equally important strand is adversarial. You check whether tasks are solvable at all as written, hunt for ambiguity a grader would trip on, and deliberately solve the same task with a different but valid package or approach — `natbib` versus `biblatex`, `amsmath` alignment environments versus a hand-rolled array — to see whether the automated compilation verifier wrongly rejects a correct answer. Those findings go back as written bug reports, and the quality of the report is a large part of what you are judged on.

What the screen looks for

  • Demonstrated hands-on depth, not familiarity. Expect follow-ups on specific error classes: undefined control sequences, missing `.sty` dependencies, `Overfull \hbox` warnings, BibTeX key mismatches, and what you actually do in each case.
  • Bibliography tooling. Zotero or BibTeX experience is listed as preferred, and it comes up: entry type selection, field normalisation, export hygiene.
  • Judgment about correctness. Whether you can distinguish "compiles" from "correct" and articulate when two different LaTeX solutions should both be accepted.
  • Written English. Bug reports are the deliverable for a meaningful share of tasks.

Logistics

Fully remote and asynchronous, structured as pay-per-task rather than an hourly contract. Average handling time is stated as 60 minutes per task, so throughput planning matters — a task you drift on for three hours still pays as one task. The engagement runs five weeks with possible extension, and the start date is immediate. Pay is not disclosed in the posting; ask for the per-task rate and the expected task volume before committing time. Entry is via a Job Interest Form followed by profile review.