The degree to which we can understand and interpret intermediate computational states during a model's inference process.