Blog
Eval³ 1.1
Eval³ 1.1 gives auditors 200 turns instead of 40, tells them how many remain, and groups issues under their root cause so splitting one flaw no longer pays.
How good are models at benchmark auditing?
Introducing Eval³
We found 110 reward hacks and grading bugs in Terminal-Bench, DeepSWE, and BFCL
Introducing EnvCheck