The question is not whether machines can replace judges, but which part of judging is verification and which is discretion.

"Can AI replace judges?" is a badly formed question, because a judicial decision is not one kind of act. It combines two: a mechanical verification of legal premises and a discretionary calibration of process.

The first is checking. Does the cited authority exist, is it good law, does it stand for the proposition, does it govern in this forum and posture, do the marshaled facts establish the element under the governing standard? These questions have answers, and machines working over a verified substrate can increasingly supply them. A judge who performs the same checks by hand is doing necessary work, but not distinctively judicial work.

The second is calibration, and it is where discretion lives. Our work describes it along four axes: how willing a judge is to terminate a case early, when the filter bites (at the pleadings or at summary judgment), how much discovery and fact development the judge authorizes, and how actively the judge manages the case. Two lawful judges can differ on every axis over the same law and reach different outcomes. Not every such ruling is equally durable on appeal, so the point where discretion truly decides a case, the durable kill point, is itself something to locate.

This is decomposition, not automation. A verification engine handles the mechanical common-law part. A measurement instrument like Ariadne's Thread estimates how much the outcome still depends on judicial disposition once the law is held fixed. Where that dependence is low, the judge is analytically unnecessary to the result, though not institutionally unnecessary. Where it is high, the judge is the decision, and the machine's job is to show exactly where.