How should machines model the other side's best case and allocate burdens?

A motion is an application for a signature. What the movant files is, in substance, a draft of the court's order: a standard selected and quietly shaded, facts arranged so the standard resolves them, and authorities lined up so that granting feels like obedience to precedent rather than a choice. An opposition that answers the motion paragraph by paragraph leaves the judge holding one coherent account and a list of complaints about it. The opposition that wins hands the court a second, better draft. This program starts from that observation and asks what it would take for a machine to do adversarial work well.

Two lines of work run through it. The first is the Counter-Model Builder. An opposition system should decompose a motion into the assertions it cannot win without, track the state of each one (holds, contested, severed, conceded) against the record and the governing standard, and assemble a complete, auditable account of why the motion fails. Retrieval and drafting are inputs to that job, not the job. The same logic governs adverse authority: lawyers rarely miss the case that sinks them for lack of effort. They miss it because they searched for their rule while the other side searched for the exception. The second line is burden allocation. American law contains at least 32 distinct burden-shifting frameworks, descended from the nineteenth-century distinction between the burden of production and the burden of persuasion, and a system that does not represent who must prove what, and when the burden moves, cannot reason about a motion at all. Loper Bright Enterprises v. Raimondo (2024), which overruled Chevron, shows how suddenly that allocation can change beneath a model trained on the old law.

Why this matters: adversarial reasoning is not a feature to add once retrieval works. It is how courts actually decide, and it is where current tools are weakest. A system that models only its own side's case produces confident briefs that come apart in the reply.