When It Breaks

The rule is not a guarantee. About a third of the time it nails the best option, which also means it misses two-thirds of the time. Two failures stand out: the best option hides inside your look-around phase and gets rejected on purpose, or every later option is mediocre and you never trigger a commit. The rule wins on average, not every single run.

So treat 37% as a starting point, not a sacred number. Shade your exploring shorter when a great option is rare and you would hate to let one slip past during the look-around phase. Shade it longer when you can afford to be picky and want a sharper benchmark. The skill is knowing roughly when to switch from learning to deciding.