JudgeAI

Normative action review, AI Arbitration, and AI Lawmaking share a single computational layer and differ only in what they assemble at the end. That means a team can move from reviewing a policy to drafting the bill behind it without switching tools or rebuilding the model. The output types are a verifiable normative decision, a full arbitral award, and a complete bill, each exportable as a PDF.

At the base sits a structured model of people, roles, capabilities, and dependencies. One person can be an employee, a parent, a property owner, and a patient at the same time, and the ontology keeps those roles linked, so a change in one shows up in the others. Companies and institutions enter through the relationships between their employees, owners, customers, and citizens. That shared vocabulary is what lets the system compare two very different situations on the same footing.

For each option under consideration, JudgeAI projects a separate possible world from a shared starting state. Closing a factory, adding a transition period, or paying compensation each produce a different world with different consequences. The engine computes how those consequences propagate among connected participants and ranks the worlds, then puts the most severe threat to any participant first. This is why the team's own study reports that removing relationships between participants flipped the preferred option in 176 of 700 tested scenarios.

JudgeAI publishes its internal studies rather than only its conclusions. The evidence room lays out the tested scenarios, the alternatives examined, and the comparisons against language models, judicial decisions, and existing legislation. One study covered 3,456 scenarios and found that expanding the set of interventions produced a strictly preferred option in 128 of them. Numbers like these are the company's own measurements, and they show a research program at work rather than a finished benchmark.

Every decision carries its reasoning: which circumstances were represented, which alternatives were compared, and which computation settled the result. Because the chain stays intact, the result becomes something to investigate. A researcher can ask how the choice depends on one data point, which disagreement between sources actually matters, and whether the answer survives a changed assumption. Recalculation then shows if the decision moved and why.

The arbitration mode closes a record and produces an award with procedural history, issues for resolution, evidentiary assessment, applicable law, reasoning, and an operative part. The lawmaking mode goes further and piles up a legislative report, a drafting blueprint, strict bill text, a validation pass, and a methodological appendix. Both outputs arrive as documents a lawyer can mark up, which is the realistic use: a strong first draft, not a filing.

The public Legislative Atlas indexes regulatory material across the United Arab Emirates, Ireland, the United Kingdom, the Netherlands, the United States, and China, with RSS and JSON feeds for tracking changes. It works as a lightweight research surface for anyone who wants to see how the same policy question gets handled in different legal systems before committing to a model.