the model releases blur together, your eval set is the only honest scoreboard

another launch, another set of charts. the only number I trust anymore is how it does on my own tasks. build the eval, run every new model through it, ignore the rest.