# AI private judge PoC repo, looking for testers + feedback

**URL:** <https://forum.liberland.org/t/ai-private-judge-poc-repo-looking-for-testers-feedback/1315>\
**Category:** General\
**Created:** [February 17, 2026, 8:22am UTC](https://forum.liberland.org/t/ai-private-judge-poc-repo-looking-for-testers-feedback/1315 "2026-02-17T08:22:28Z")\
**Posts on this page:** 2\
**Page:** 1

<div class="post-metadata">

**Author:** ![J-B](https://forum.liberland.org/letter_avatar_proxy/v4/letter/j/919ad9/32.png) [@J-B](https://forum.liberland.org/u/J-B)\
**Post date:** [February 17, 2026, 8:22am UTC](https://forum.liberland.org/t/ai-private-judge-poc-repo-looking-for-testers-feedback/1315/1 "2026-02-17T08:22:28Z")

</div>

Dear Liberlanders,

I recently worked on the following question: can **AI** be judge, for a court?

Liberland **authorises explicitly** the use of private judges, so there is a legal framework to allow this. I have to admit I am not fully clear on the **Judge App** thing, but this is the concept.  
Also, as far as I understand, AI is listed as allowed.

Also, I think LLMs are now powerful enough in terms of reasoning and in terms of tooling to deliver this service (RAG, file search). Obviously, the **digital-native** legal infrastructure of Liberland also helps!

So now is the moment for a proof of concept.

In the repo [https://github.com/Dessoul/ia-court](https://github.com/Dessoul/ia-court) you will find all the requirements to run a court with 3 AI judges. It is using Claude Code out of the box, but you can run it on any LLM.

This is a bit the goal of having 3 judges that need to agree together: they can be different LLMs, or LLMs with different roles (one can be systematically skeptical, the devil’s advocate, etc.).

This is an iterative process in which the judges have to finally agree on the verdict. They can ask questions to the different parties of the trial.

In the repo you will also find some propositions of cases, that are synthetic, AI-made of course. Side note: this is very funny how AI created fake names like DanubTech and stuff 🙂

They are business-related cases, as it seems logical to me that they are the best candidates for first tests.

I would be happy to help you test it if you need any help; or provide more information on how it is supposed to work and the detail of the process to coordinate the different judges while giving them autonomy.

I am also very interested to get your feedback to improve this tool, to move from PoC to MVP. **If you have a real low-stakes dispute where both parties consent, I’d love to try it too.**

---

<div class="post-metadata">

**Author:** ![Jae\_Kim](https://forum.liberland.org/letter_avatar_proxy/v4/letter/j/b9e5f3/32.png) [@Jae\_Kim](https://forum.liberland.org/u/Jae_Kim)\
**Post date:** [October 30, 2026, 2:33am UTC](https://forum.liberland.org/t/ai-private-judge-poc-repo-looking-for-testers-feedback/1315/2 "2026-10-30T02:33:05Z")

</div>

Thank you for building this and putting it out in the open. I read through the protocol, and several choices stand out:

- The clerk is deliberately kept weak and can’t frame the issues.
- Round 1 has no vote, so judges don’t lock onto an early position.
- Corrected evidence supersedes the original instead of replacing it.
- Each judge’s questions are capped and passed on word for word.

Your finding that the skeptical judge did best in testing makes sense. Holding off on a judgment tends to produce better analysis, in people as well as models.

Since you asked for feedback, here are three suggestions. Each adds a heading or two to the filing templates and leaves the process itself alone. I’ve tried to explain why I think each one matters, since the reasons are what you’d want to check.

**1. Ask for the record made before the dispute**

The filings start once the dispute already exists. By then each side knows which side it’s on, and that changes how people read the same facts. This isn’t dishonesty; it happens even to people who are trying to be fair.

Researchers have measured it. In a well-known series of experiments by Linda Babcock and George Loewenstein, people read the same case file and then predicted what a judge would award. Those told in advance they were the plaintiff predicted much higher awards than those told they were the defendant. The bias disappeared when people learned their role only after reading the facts, and those pairs settled more readily ([Babcock, Loewenstein & Issacharoff, _Law & Social Inquiry_, 1997](https://www.cmu.edu/dietrich/sds/docs/loewenstein/CreatingConvergence.pdf)).

Real parties always know their role. But they usually left evidence from before they did: the brief, the quote, the messages where they agreed on scope and deadlines. Those were written before anyone had a side to defend.

Take a typical case. A client says they ordered a full website redesign of ten pages. The designer says it was a refresh of five. Each “facts and timeline” section will be honest and different. The original quote and the email accepting it will usually settle the question in a minute.

Since the Judicial Process Law admits all evidence, the question is how much weight it gets. A heading such as “What was agreed at the start, and where it is recorded” would point the judges to the evidence least shaped by the dispute. It should also save some of the three questions per round that judges would otherwise spend reconstructing what was agreed. Over time it would reward people who write down what they agree before they need it.

**2. Ask where every number comes from**

The claimant’s template asks for a “specific amount,” but not how it was reached.

That matters for two reasons. First, a total can’t be checked without its parts. Second, without the parts the judges are pushed toward all or nothing: award the claimed figure, or award nothing.

Suppose a claim for €4,800 is shown as 40 hours × €120, and the defendant accepts the rate but disputes 12 of the hours. The judges can then decide the 12 hours on the evidence and award €3,360 if the defendant is right. Without the breakdown, they’re weighing €4,800 against zero.

The same applies to the proposed order. If every figure in it traces to an evidence ID or a clause, anyone reviewing the decision, including an appeal court, can follow the arithmetic. They don’t have to trust it.

**3. Invite each party to state the strongest point against its own case**

This sounds odd at first, but it’s the suggestion with the most direct evidence behind it. In the same 1997 study, half the negotiators were simply asked to list the weaknesses in their own case before negotiating. Nothing else changed. The results:

- **Settlement:** pairs settled 96% of the time, against 65% without the exercise.
- **Bias:** the gap between the two sides’ predictions of the judge’s award fell from about $21,800 to about $4,700.
- **Cost and speed:** they settled in roughly half the time, at roughly half the legal cost.

It does for the parties what your skeptical judge does for the panel.

In your process it would also narrow Round 1. Whatever a party concedes moves straight into “uncontested,” so the judges can spend their limited questions on what’s really in dispute. I’d keep the heading optional and say clearly that conceding a point is never held against a party. Otherwise nobody will fill it in honestly.

**Why these three together**

These habits come from a short method I wrote up for preventing and resolving disputes:

- register each side’s goals separately
- enter evidence with dates
- cite the source of every number
- write the case against yourself
- correct openly
- date who checks what

It’s free to use, if helpful: [The RECORD Method: A Complete Guide | Zenodo](https://doi.org/10.5281/zenodo.23327427)

Your protocol already does the last three steps well: superseding evidence, dated rounds, and a public verdict with appeal information. The suggestions above add the first three at the point where cases come in.

**A small test**

Take one of your synthetic cases and run it twice with the same models: once with the current templates, and once with the three headings added. Because model output varies, three runs of each would be better.

Then compare a few measures, agreed in advance:

- how many rounds it takes to converge
- how many questions the judges need to ask
- how many facts end up marked “contested” in Round 1
- whether every figure in the final order can be traced to its source

If the headings make no difference, that’s worth knowing, and they’re easy to drop. If they help, it’s a cheap improvement before you move to real cases.

**The bigger opportunity**

Projects like yours point to something larger. If it’s built carefully, I believe Liberland could become the jurisdiction of choice for international business disputes, especially smaller cross-border ones between freelancers and small to mid-size contractors.

These people have almost nowhere to go today. Take a designer in one country owed €3,000 by a client in another. Suing abroad means foreign courts, foreign lawyers and a foreign language. Commercial arbitration is built for much larger sums, and its fees can exceed the claim itself. Platform dispute systems only cover work done on that platform. So most of these disputes are simply written off. The one who was wronged absorbs the loss, and both sides learn not to work across borders again.

Liberland already has several of the pieces this needs:

- a legal framework that explicitly allows private judges
- proceedings designed to run online
- a judicial marketplace in which parties choose who decides their case
- people like you building tools for it

And because the Constitution and the main laws are still being drafted, a few targeted edits now could make Liberland a much friendlier place to bring these cases. Examples include clear rules on consent to a private court’s jurisdiction, how settlements and awards are enforced, and how mediation works. Changing those things later, once the laws are settled, will be much harder.

The obvious objection is enforcement: Liberland isn’t a party to the New York Convention, so its awards can’t yet count on enforcement abroad. For small claims, though, much of that can be handled up front. If the parties agree at the start that payment sits in escrow and is released according to the decision, enforcement happens without any foreign court involved.

Over time, every case resolved fairly would help build Liberland’s economy and its reputation. Each one would be a working relationship that held up instead of falling apart, and repeated across many cases, that’s the kind of track record that wins recognition from the international community.

I’ve made suggestions along these lines in comments on the draft laws, most directly on the courts, on enforcing agreements, and on companies:

- [Draft Constitution](https://docs.google.com/document/d/1oEJWvy6c7Xf2wlG6RhbFzrdC-c_06bJ-QHROo_kYlkk/edit)
- [Judicial Process Act](https://docs.google.com/document/d/1IG_aj6-VBOKunAymDDi6NspiuX0olZjO1oeSNOV3z-Q/edit)
- [Company Law](https://docs.google.com/document/d/1ZvO27riEu2ReidW2J7RLe92R2ZjLMKRYrbDt-VZYdIw/edit)

I’d welcome feedback and comments on any of them, from you or anyone else here. Counter-arguments and devil’s-advocate positions are especially welcome. A proposal that hasn’t survived its strongest objections isn’t ready.