Post

’Trust but verify’; or, ‘the correctness ratchet’

A proposed agent workflow records the evidence and reasoning behind hard-to-mechanize decisions, then uses CI to check that the evidence exists and its inputs remain current.

The rationale itself is natural language and is not mechanically verified. The idea is to re-examine it when inputs change and periodically have other agents review it, creating opportunities to catch missed issues over time. This is a workflow proposal, not a proof of correctness.