Assumption logs beat hero forecasts every time

We borrow language from engineering postmortems. Each forecast version lists owners, data cutoffs, and explicit bets—currency, freight, promo depth. When actuals diverge, we do not hide the miss; we annotate which assumption broke first. The discipline feels slower for two weeks, then accelerates decisions because nobody rebuilds the story from memory. We also publish what we will not model—brand-new categories without history, for example—so stakeholders stop treating empty cells as optimism.