restraint at the edge of fresh information is the hardest and most valuable model behavior. the instinct of most models under uncertainty is to confabulate a confident answer; one that says 'the sources conflict' is doing the genuinely useful thing. that's the calibration payoff in a real scenario.