Cheap Judgment Still Has to Earn Authority

A narrow decision model does not need to write an explanation, hold a conversation, or invent a plan. It can choose from a fixed set, score a known condition, or say how likely a statement is to be true. That smaller job matters because factories make the same kinds of decisions constantly: match or reject, route or escalate, accept or inspect.

When each judgment costs almost nothing, the architecture changes. You no longer need to sample a tiny part of the queue and guess what the rest looks like. The judge can observe every case, record its answer, and leave the existing process untouched. Cheap judgment buys coverage before it buys control.

That is the first move we would make. Put the model beside the route, not in charge of it. Let it write a shadow decision with its version, inputs, confidence, and chosen class. The factory gets a complete map of what the model would have done without pretending that a plausible answer has earned production authority.

Advisory mode is safe, but it can become a comfortable graveyard. Teams add a score to a table, congratulate themselves for being cautious, and never build the path that turns evidence into a decision. The model keeps spending money to produce a column nobody uses. Nothing improves, and no human work disappears.

A useful shadow run starts with a promotion rule. Name the cases that will be reviewed, the outcome that counts as correct, the kinds of mistakes that matter, and the amount of evidence required before one narrow class may proceed automatically. If nobody can say what would earn promotion, the trial is research theater.

The exit cannot be one average accuracy number. A wrong recommendation on a reversible catalogue match is not the same event as a wrong recommendation on access, money, or deletion. Authority should expand by consequence class. The harmless, observable route can move first. The irreversible route may never belong to the judge at all.

A judge that can express uncertainty gives the factory another useful output: not this one. But an uncertain score sitting in a log is not a control. The system needs a real abstention route that gathers missing context, asks a stronger evaluator, or sends the case to a person who owns the consequence.

That route should preserve the reason for the stop. Was required data missing? Did two signals conflict? Did the input fall outside the patterns the model had seen? Did the score land near a boundary? Those are different failures, and collapsing them into a human queue throws away the information needed to shrink that queue later.

People should handle the cases that can still teach the system something. They should not repeatedly clear the same ambiguity because the factory forgot the last answer. Every override needs to become a labelled example, a changed threshold, a tighter deterministic check, or a newly split class. Otherwise advisory judgment only creates a second backlog beside the first.

Not every judgment belongs in a model just because the model is cheap. Exact sizes, dates, permissions, balances, and identity checks should stay in deterministic code when the system can compare them directly. Asking a probabilistic judge to rediscover a hard contradiction makes the route less reliable for no benefit.

Use the judge where interpretation is actually required: messy names, incomplete descriptions, competing categories, or language that does not fit a clean rule. Then surround that judgment with facts the model cannot negotiate. The factory should reject an impossible size match before the model gets a vote, and it should block a forbidden action after the model speaks.

This division is what makes a specialized model valuable. It does not replace the control system. It fills the narrow gap between rules that can be computed and cases that still need judgment. The model handles ambiguity; the harness keeps authority.

Start with one high-volume queue whose result can be checked later. Run the judge silently. Sample its confident answers, inspect every costly miss, and connect its decisions to real outcomes. Then grant one bounded action to one proven class. Watch it again. Revoke that authority automatically when the inputs drift or the error budget breaks.

That loop is slower to describe than plugging a model into a workflow, but it is the part that compounds. Once promotion and revocation are machinery, the next judge does not begin with a debate about trust. It enters the same probation, faces the same evidence, and receives only the authority it earns.

Cheap decision models will eliminate a lot of routine triage and review work. They should. People are an expensive way to repeat a stable boundary thousands of times. Human judgment belongs one level higher: choosing the consequences, defining the promotion rule, and owning the cases the machinery cannot yet reduce. The low price opens the door. The factory decides how far the judge gets to walk through it.