Goodhart Labs does alignment work on training data and environments. We do independent QA and evals for the reinforcement learning environments used to train frontier models.
The integrity of training data and environments is alignment-critical. Our work covers QA for the producers of those environments, evals on the gap between target and trained behavior, and alignment research.
For model labs.
QA and evals on the RL environments you train on. We look for reward hacking, spec drift, and integrity issues.
For RL data companies.
Pre-delivery QA on environments before they ship to labs.
Careers.
Hiring ML, alignment, and security engineers. Send work or a resume to hello@goodhartlabs.com.