A model answered our question. It would not tell us why. That refusal is the entire case for governed AI.
Today we added a new frontier model to SAFi, pointed one of our agents at it, and watched every single turn fail.
The response was always the same: the language model returned an empty response. The error blamed the API key. The key was fine. Other models on that same key, answering that same question, worked without complaint. So we looked closer.
The call to the vendor had succeeded. HTTP 200, a clean response, no error anywhere in the transport layer. But the response carried no answer. The vendor’s own field for why the model stopped held a single word: refusal. Zero content. Zero output tokens. The model had been asked a question, had decided not to answer, and had said so through the only channel it has for saying so.
For a harmless question, that was surprising. So we narrowed it down. Same model, same instructions, same question, with one thing stripped out: the instruction SAFi attaches to every drafting request, the one that asks the model to explain its own reasoning. With that instruction removed, the model answered normally. With that instruction present, and nothing else changed, it refused. Every time.
