Role Description & Scope
**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.
1. **Design the challenge** — identify the hidden basin and the forcing chain that makes one conclusion inevitable.
2. **Build the world** — assemble a data room from real, source-grounded data. Every decisive value traces to a real source.
3. **Write the grader** — a grading document that scores a submitted memo on substance, not style or method.
4. **Validate it** — run verification and calibration before submission.
5. **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.
**Commitment**
Minimum 20 hours per week, up to 40.
**Tooling**
Work happens in a browser-based studio plus Claude Code. A Claude Max subscription is required and is fully reimbursed.