Judgement does not end at the choice, it runs the whole line

Last week I split a boundary into four things: a constraint, a decision, an invariant, a check. The constraint is the rule, the decision applies it, the invariant is what others rely on, the check enforces it.

It is tempting to read the last three as mechanical. They are not. The decision spawns the invariant, and the invariant selects the check, but each handoff carries a judgement: does the invariant faithfully say what the decision obliges, and does the check enforce it rather than a convenient proxy?

So “can AI make the decision” is malformed. Beneath the decision sit the judgements that carry it down. Judgement does not end at the choice. It runs the whole line.

And the line reshapes as it goes. Choose microservices and you open a platform choice that was not there before. Choose a monolith and you close it to one. Same place, opposite shaping.

This is where the work stays with people. Enforcement can be handed to a machine: a check, once written, just runs. What cannot be handed over is the judgement that the check stays faithful to the invariant, and the invariant to the decision. The earlier the judgement, the more it has to trace to someone answerable. You hand over the running. You keep the faithfulness.