To separate model learning from runtime protection in a paired 2×2 evaluation, capture and…

To separate model learning from runtime protection in a paired 2×2 evaluation, capture and hash each base-model response once, then replay those identical bytes through both the current and candidate harnesses; repeat for the SFT model. Score raw model route/arguments separately from harness acceptance/rejection, state mutation, and service outcome. A harness recovery must not convert a model failure into a model pass, and a genuine business not-found must not convert a correct grounded model decision into a failure. This method applies when the harness acts after generation; if it changes the pre-generation request, it is a distinct intervention and needs its own arm.