Multilateral Autonomous Legal Reasoning Engine for Customary International Law and Confidential Arbitration

Al systems that merely retrieve legal texts do not become competent in law when the target domain is uncodified, non-monotonic, and strategically hidden. We introduce MARE (Figure 1), a Multilateral Autonomous Legal Reasoning Engine for two international-law use cases: (i) International Court of Justice identification of customary international law from state practice, opinio juris, advisory opinions, and defeaters; and (ii) London Court of International Arbitration strategy under confidentiality, selection bias, and hidden arbitral analogues. The first module compiles legal evidence into a weighted epistemic-defeasible argumentation graph and computes norm survivorship under grounded semantics. The second module models arbitration as a hidden-type Bayesian game/POMDP and computes policies by belief updating and regret minimization rather than supervised prediction over public awards. We report three self-contained prototype experiments: a synthetic ICJ custom stress test, an advisory-opinion influence graph, and an LCIA hidden-type arbitration game. The experiments are not presented as measurements of tribunal outcomes; rather, they are validation probes for the architecture's formal failure modes. MARE improves non-monotonic custom classification from 69.7% to 89.0%, improves advisory influence NDCG@8 over citation baselines, and raises hidden-type worst-case arbitration value from 0.418 to 0.512 while reducing best-response vulnerability. The contribution is an architectural blueprint for building and evaluating legal reasoning systems whose first-class objects are norms, defeaters, beliefs, and strategic uncertainty.