verityResearch freezes the verdict before the experiment

Field notes

The project uses a fixed sequence: define both arms, commit the interpretation and pass/fail rules, then read the evaluation output. Its sole listed study reports two evaluation sets—a held-out set of 106 prompts and an adversarial probe of 43—with answer rates as well as tool-call rates. The homepage links to a full write-up and GitHub code, and identifies the license as Apache-2.0.