Worked proofs, not brochures
Case studies
Each of these states an outcome up front, generates data that satisfies it, and then checks the result with something that shares no code with the generator. Every figure was measured on an actual run, and the seed and engine version are printed next to it so you can rebuild the file and disagree with us.
Anti-money laundering
Declare the laundering rate, the typology mix and the hard negatives, then prove nothing else in the file looks like a crime.
Financial crime teams, model validators, and anyone benchmarking a detector
Verified eval databases
Text-to-SQL benchmarks are wrong about half the time. Declare the answer first and the answer key cannot be wrong by construction.
Anyone training or evaluating a data agent, and anyone who needs a reward signal they can trust
Other proofs on this site
These predate the case study format and are big enough to stand on their own, but they answer the same question: does the thing hold up when somebody checks?
A conformance suite of SQL assertions run against the engine's own output, including the ones it is known to fail.
Real open-source dbt pipelines inverse-solved to declared outcomes, bugs in the solver included.
Timings and memory on real hardware, with the envelope stated rather than the best run.
Want one built for your domain? Tell us what you would need proved.