Case studies
Work we've done for buyers: what they asked for, how we tested it, what we found and its limits. We don't name the buyers.
- Contamination-proof coding environments at 300-600 accepted tasks per weekA sustained supply line of graded coding environments for RL training, where each task ships with executable proof that a model cannot pass it by reproducing a library it saw in pretraining.
- Testing a compute panel for equity alpha, and publishing the nullA four-year cross-provider GPU rental panel built to test whether compute input costs predict the equity returns of the companies buying them. The directional result was null, and it was published.
- Four integrity tests a sentiment panel has to pass before a backtest means anythingA fund that had already rejected two sentiment vendors ran a four-point integrity audit against eighteen years of human-labeled retail sentiment before licensing any of it.