Guides and sensors

A harness has two kinds of part, and teams that build only one kind get a predictable failure.

Each can be computational or inferential. A computational check is deterministic, instant and nearly free — does this parse, is this field present, was this tool called. An inferential check asks another model, which is semantic, slow and probabilistic. Both are real; they are not interchangeable.

• Guides only, with no sensors, is unverified theatre: a long document of instructions nobody has confirmed the agent follows.
• Sensors only, with no guides, means catching the same mistake repeatedly instead of preventing it.
• Put the cheap computational checks everywhere, because they cost nothing to run on everything.
• Spend the inferential checks by criticality, not uniformly — the reply that moves money deserves a model-graded check; the status lookup does not.

The useful habit is what to do when the same error recurs. The reflex is to edit the prompt. The discipline is to ask which part of the harness should have stopped it — a guide that was missing, or a sensor that did not look — and strengthen that instead. One fixes an instance; the other fixes a class.