LatchBio researcher @arjunomics reveals Grok's refusals often come from the model itself while Fable and GPT-5.6 Sol rely more on external safety layers: "There might be input classifiers where when a user puts in a query, the model initially might say, this is a bad query, we can't respond to this. There might also be output classifiers, so after the model outputs a query, you might have an external classifier that says, no, don't show this query, put it in an API block. And there are also…