Maestro

Checks and re-routing

Every result is checked by something that did not produce it. When a check fails, only that task moves to a stronger model.

What counts as a check

A check is a test, a review or a rule that a result must pass. It is always run by something that did not produce the result:

Kind of workChecked by
CodeTests, plus a review by a different model
Copy and contentA review
Fixes in a runTests, a scan for secrets, file limits and cost, then a review
Anything with a ruleThe rule itself, such as where data may go or how much a task may spend

When a check fails

The task is re-routed: it goes to the next model up and tries again. Only the failed task moves. The rest of the work stays where it is, on the models that did it well.

Steps column: Sonnet wrote the tests, one failed, and the fix went to Opus.
Sonnet wrote the tests. One failed. The fix went to Opus; the files stayed with DeepSeek.

You pay the higher price for that one fix, not the whole job. In the example store build, seventy-nine of 3,559 tasks failed a check and were re-routed, and their retries are inside the $29.10 total.

In the record

Every retry is in the record: the task, the check it failed, the model it moved to, and what the retry cost. In a run, a fix that fails three tries is marked failed and kept in the record, so nothing that was tried is lost.

Was this page helpful?