Five parts govern one task

01

A defined task

The business already has the task. People have a name for it, perform it, and carry its rules and exceptions in their heads. The first step is to make that implicit knowledge observable, not to invent new work.

A defined task has observable input and output. Its executions must also be the same kind of thing, so their results combine into a meaningful measure. Store hours and thousand-dollar refunds do not belong in one support pass rate.

02

Independent grading

You need to capture how the task is being performed, and something needs to grade that work. The grader is separate from the worker doing the task.

Grading can come from policy checks, classifiers, model judges, human review, or a combination. The right approach depends on what the task requires and the cost of a wrong verdict.

03

Evidence retained across executions

This is not a one-time check. Evaluator results accumulate across executions so a decision rests on sustained performance rather than a convincing example.

Keep relevant operating costs alongside those judgments. Behavioral evidence establishes whether the configuration meets the task's bar; economic evidence shows what acceptable outcomes cost, including evaluation, review, retries, and escalation where relevant.

The record remains available after a decision. Continued review adds the difficult cases back to it, helping the evaluator improve and revealing what may be safe to automate next.

04

Promotion and demotion rules

A promotion rule states what evidence makes a task eligible for less oversight. A support task might require enough qualifying cases above a success threshold with no policy violations. Passing that rule makes the change eligible; approval or the declared automatic rule still has to apply it.

The same evidence may justify keeping or replacing the operating configuration without changing authority. Qualification and promotion are separate decisions; a cheaper worker must still satisfy the behavioral bar.

Demotion rules state when tighter oversight returns. They can respond to a measured trend or a single severe failure. Operators must also be able to intervene without waiting for a statistical window to close.

05

A real fallback

Fallback is what happens when you pull the ripcord, and the ripcord has to be connected to something. The safer mode must remain operational, whether that means routing support work to a person or pausing it.

Exercise that path during routine operation. Restoring oversight cannot depend on an emergency procedure nobody has tested.