About the role
This is an evaluation framework for frontier AI models in drug research and development. Domain experts write realistic research problems wrapped around messy evidence worlds. The model works offline in a sandbox with open-source scientific tooling, receiving incomplete, indirect and sometimes misleading data, and must reach a defensible conclusion. The answer is present in the evidence but never stated.
Your job is to build a research problem so realistic and well-constructed that the model has to genuinely reason to solve it — it can't pattern-match or look the answer up.
What you'll deliver
Three things per task: the prompt, the data room, and the grading document.
- Design the challenge — identify the hidden basin and the forcing chain that makes one conclusion inevitable.
- Build the world — assemble a data room from real, source-grounded data. Every decisive value traces to a real source.
- Write the grader — a grading document that scores a submitted memo on substance, not style or method.
- Validate it — run verification and calibration before submission.
- Revise it — you own the rework if a reviewer requests changes.
You are the final authority on every scientific and difficulty question in your task.
Who we're looking for
- Working expertise in mechanistic enzymology, structure-guided peptide or antibody design, fragment-based discovery, or structure-based medicinal chemistry — PhD or equivalent industry track record.
- You can pose a problem whose answer you can defend rigorously, but which a strong solver cannot shortcut.
- Comfortable judging reasoning quality, not just final answers.