How it works
Second look and verdicts
Every report gets a second look before you read it. The verdict grades the report against its own evidence. It never grades a token.
The verdicts
| Verdict | Means |
|---|---|
| Holds up | The answer and findings follow from the steps, and the method matches the plan. |
| Doesn't hold up | A claim isn't supported, a number is implausible, the method doesn't match the plan, or groups were compared unfairly. |
| Unclear | The evidence is too thin to say, or the judges disagree without either finding a problem. |
"Holds up" never means safe. A report can hold up and describe a token whose creator owns 92% of the supply. The verdict is about whether the report is right, not whether the token is good.
Layer 1: rules in code
Before any model looks at it, code checks that every number came from a tool or the question, that there's no advice, prediction or labelling, and that every address came from the tools. A report that fails is sent back to the agent; if it still fails after two tries it's filed as "doesn't hold up".
Layer 2: two judges
Two models from different families read the plan, the report and the evidence and grade it. A report holds up only if every judge says so; if either finds a real problem, it doesn't hold up.
A judge never grades a model from its own family, because it tends to be lenient. That report is graded by the other judge alone. Today's judges are GPT-5.4 mini and Claude Haiku 5.5.
What judges look for
- Claims that aren't in the evidence, or contradict it.
- Implausible numbers (trillions of dollars of volume for tokens a few days old).
- A method that drifted from the plan's definitions without saying so.
- Unfair comparisons: one group filtered one way, the other group another way.
Judges make mistakes
Judges are models too. In our own tests one judge misread what a tool covers, and another missed a pair of numbers that couldn't both be true. That's why the verdict comes with each judge's explanation, and why more layers are planned: an independent verifier (Jev) and, for public reports, human review.