How do you respond to a colleague who suggests replacing CrAM with CAG after a security failure?

Before swapping to CAG, you must isolate the failure point. CrAM does not compute credibility; it only acts on the provided scores. If a document was correctly scored low-credibility but still influenced the output, the failure could be due to renormalization—where freed attention mass is redistributed into other unreliable tokens—or because the gullible-head selection is stale or task-specific.

Swapping to CAG changes where the credibility judgment is stored, but it does not eliminate the underlying risk. If the credibility judgment itself was flawed, or if the system is not properly calibrated for the current domain, CAG will suffer from the same fundamental issues as CrAM.