What is judgment?

Code grades itself. Write the function, run the tests: it passes or it fails, in seconds, and nobody's opinion enters into it. Ground truth in code is objective and immediate. That is why machines got good at code first.

Some work gets its verdict eventually. The surgery has an outcome. The trial ends in a ruling. The feedback is slow and noisy, but it arrives, and over a career it teaches.

Then there is the work that gets no verdict at all. A contract is signed, a structure chosen over the nine others the rules allowed, and then silence, for years. No ruling ever comes. The silence sounds the same whether the call was right or wrong.

In work like that there is no test suite. There is judgment. Ask a lawyer why they struck the clause and the answer will use that word: in my judgment, the risk was not worth carrying. Press, and you will get reasons. The reasons will be good ones. Reasons were never the problem.

You have watched what happens when two sets of good reasons meet. A negotiation between sophisticated lawyers, each with a defensible position, neither able to prove the other wrong, because there is no test that settles it. It goes a few rounds, and then the point goes to the clients to settle as a business decision. The most heavily lawyered conversations often end with the lawyers stepping aside.

That was often where it ended. The most experienced person in the room was the ceiling. Events tested the judgment sometimes, years later. Often, nothing ever did.

Frontier models change that without solving it. A model has read more law than any lawyer will in a lifetime and can hold more of it in view while working through a single problem. It can challenge a lawyer's judgment in ways another lawyer often cannot.

It can also be wrong. That is what makes the encounter useful. Neither side gets to be ground truth.

That disagreement is useful because it tells you where to look. A model can expose a weakness in a lawyer's judgment. A lawyer can expose a failure in the model. Neither tells you, by itself, what the right answer is.

That is the harder problem. Legal practice produces judgments and reasons for them. Training requires knowing which of those judgments should become signal. The disagreement has to be resolved, and the result validated, before either side deserves to teach the other.

Skip that work and the machine learns from the proxies instead. Confidence. Explanation. Agreement. Credentials. The surface of judgment without establishing whether the judgment was right. A supercomputer being taught to swagger.

Decoding judgment is the work of turning what legal practice leaves unresolved into training signal whose quality can be established.