RL environment and verifier authors.
People who build the places where models learn. An environment is only as good as its verifier, so they think about shortcuts as hard as they think about tasks.
What these contributors do.
Tasks with a clear goal, realistic state and a single defensible answer or an outcome a program can check.
Checks that reward the right behaviour and reject near misses, written so they hold up when a model probes them.
Trying to pass each task without solving it, and closing every shortcut that works.
Running current models against tasks and tuning them so they measure something, not too easy and not impossible.
Evaluated on the work itself.
Contributors build a small task and its verifier in their own area. A reviewer then tries to break the verifier and reads how they handled shortcuts. This is the work Blackdrome does in its own research.
Who we are looking for.
Engineers and researchers who have built benchmarks, test harnesses, simulators or graders, and who like finding the hole in a check. Joining is free, and we never charge contributors a fee.
Need RL Environments experts?
Tell us what you need in a line or two. A person on our team reads every request and replies by email.
Other areas: Coding & Software Engineering · Agent Evaluation · Technical Review