How we judge

The jokes are for fun. The judging is done carefully. Here is exactly what happens to a case.

The steps

1. A private safety question. Each of you is asked, alone, whether you are ever afraid of your partner. If the answer is yes, or if a statement describes violence, threats or control, the case stops. Those situations get support information, never a score.
2. A clerk rewrites each side. An AI turns each statement into a short, neutral case file. Names, he, she and every other gendered word are removed, along with insults. Anything that tries to give the judges orders is dropped.
3. Private follow-up questions. In the full court, each of you gets two questions. Yours are written by an AI that has only seen your own side, so a question cannot give away what your partner said.
4. The two sides are sorted, not stacked. AI judges who read two stories one after the other tend to side with whichever they read last. We measured this on our own judges. So a clerk first sorts both sides into facts you both agree on and claims only one of you makes. The judges read the shared facts first.
5. A panel of judges, in both orders. Two AI judges score the case separately. One sees Partner 1 as "Person A", the other sees Partner 2 as "Person A". If they clearly agree, that is the verdict. If they differ by more than 10 points, read the case differently, or either one is close to 50/50, two more judges are called in. If flipping the labels moves the result a lot, the verdict says so.
6. The maths is done in code. The percentage is the average of the judges. If they are split, or the result is within a few points of even, the verdict is "too close to call". If most judges see a clash of needs with nobody at fault, the verdict is "you are both right" and there is no percentage.
7. A writer puts it into words. A last AI writes the verdict and the jokes. It is told the result and cannot change it.

What we cannot promise

• The judges are AI. They only know what the two of you wrote, and they can be wrong.
• The percentage is a judgment, not a measurement. Run the same case twice and it can move by several points.
• The free quick ruling is one judge taking one look, with the two of you listed in coin-flip order. The full court can disagree with it.
• Our fairness checks so far are a handful of test cases, including the same case with the seats swapped. A full published test comes before the public launch.
• This is not a real court, and it is not therapy or legal advice. Nothing here is legally binding.

Your words

• Your partner never sees what you wrote. They see the verdict, which describes what happened.
• What you typed is replaced by the neutral case file as soon as both sides are in. The case files and your answers are deleted the moment the verdict is ready.
• Unfinished cases are deleted after 30 days. Either of you can close a case at any time, which deletes it straight away.
• There are no accounts. Each of you is recognised by a private cookie in the browser you used.
• To judge a case, the text is sent to Anthropic's Claude AI. It is not used for advertising.

Who pays for this

Right now, we do. Every full verdict costs real money, because several AI judges read each case. There are no ads and no subscription. If a verdict helped, you can buy the judge a coffee. It is optional, it buys nothing extra, and it never changes a verdict.

The research the judges may cite

A verdict can only quote findings from this list. The page prints the source itself, so the AI cannot make one up. We say how strong each finding is, including when it is weak.