Anthropic explains Claude degradation after harness and instruction changes highlights a reliability issue many enterprise AI teams quietly worry about: model quality can move backward when surroundin...