What the work involves
This is asset-building, not prose writing. Turing constructs benchmark environments in which AI agents operate real desktop applications, and this role builds the LaTeX side of that: templates that map what a TeX toolchain can actually do, plus the base material tasks run on. Day to day you will be producing multi-file document repositories with realistic cross-references and includes, BibTeX or BibLaTeX citation databases, custom class and style packages, and TikZ vector figures — all internally consistent, because tasks are generated from these assets and any contradiction propagates.
You will also complete reference solutions on record. That serves two purposes: proof the task is humanly achievable, and a canonical answer for what correct means. You are not writing the scoring code — engineers implement the checks — but you define the correctness criteria they encode. That means articulating why a compiled document, a bibliography style, or a citation graph is right, in terms someone who does not typeset for a living can turn into a test.
What the screen looks for
- Verifiable depth in LaTeX: package internals, `.cls` versus `.sty` boundaries, `\cite` resolution behaviour, build-order effects with BibTeX and `latexmk`.
- Evidence you have shipped real academic or technical publications — multi-author, multi-file, with a style guide or publisher constraint to satisfy.
- Named tool fluency, not familiarity: Zotero, TeXStudio, Overleaf/ShareLaTeX, with at least two used seriously.
- Judgment about correctness and edge cases — where a document can compile cleanly and still be wrong.
Logistics
Remote and async, with at least four hours of daily overlap with US Pacific time. Stated commitment is 40 hours per week on a five-week contractor assignment with immediate onboarding, no medical or paid leave. Pay is undisclosed on this listing; Turing describes compensation as competitive and typically sets rates during screening based on domain and experience.