What the work involves

You write single-turn prompts in your own area of chemical defense and label each one at one of three levels: benign, dual-use, or adversarial. Then you evaluate what the model returns against a written policy standard and decide whether it was handled correctly — including the cases where over-refusal is the failure. Each item ships with a reference answer: what a correct response should have said, and the technical reasoning behind it. Expect the writing to take as long as the chemistry.

The domain sits on the dual-use line more than most. Toxicology that fixes a countermeasure dose also describes an exposure; decontamination efficacy data that protects responders also indicates what defeats it. A model that refuses a routine PPE, antidote, or detection-limit question has failed as clearly as one that answers a request it should have declined. That judgment is the deliverable.

What the platform screens for

  • Institutional defensive experience — countermeasures, protection, detection, verification, destruction, or medical management — rather than general chemistry.
  • Whether you can articulate why a specific prompt crosses the line, not just that it does. Follow-ups will push on borderline cases.
  • Demonstrated technical writing: publications, agency reports, expert witness work. Include a sample or link.
  • A clean answer on classified, export-controlled, NDA-bound, or prepublication-review material. Holding obligations is not disqualifying; the work is scoped around them. Concealing them is.
  • Comfort reading and writing about misuse scenarios for sustained stretches. Experts are briefed in advance and can pause or withdraw without penalty.

Logistics

Fully remote and asynchronous, paid per completed task in the $65–75 band as observed on the platform. Task volume varies with batch cadence rather than arriving as a steady weekly load, so most contributors treat it as supplementary to existing work. Calibration materials and the policy standard are provided; disagreement with a calibration call is expected to be argued in writing, not silently absorbed.