Often, evals are very disconnected from actual utility. For example, we had an eval for a while that measured 'writing style'. Basically, how well do we prevent AI slop in writing output? We maxed out the eval, put the model in prod, and users hated it.