the-promo-impact-checklisted

> **A cliff hides the cases worth catching.** A single hard multiple or fixed percentage, applied to a > population whose own spread it ignores, fires constantly on naturally volatile units and stays silent > on the ones that matter. Two consequences: > > - **Use a band, not a cliff.** Between roughly 1.5x and 2x the norm is *slipping* and gets reported > as a watch item; past 2x is *breached*. The highest-value case is routinely the one sitting at 1.6x, > trending, and invisible to a 2x test. > - **Compare each unit against its own variability, not one global number.** A metric that swings 30% > week to week and one that swings 3% cannot share a threshold: the first alarms every week and the > second never alarms at all. Where enough history exists, set the band from the unit's own trailing > spread and say you did. Where it does not, use the fixed rule and **say it is a fallback**. > - **Report the direction of travel alongside the level.** A unit at 1.4x and rising and a unit at 1.9x > and falling need opposite responses, and a level-only test cannot tell them apart. # The Promo Impact Check Take a promotion that already ran and measure what it actually did to profit, not just to the revenue chart during the sale. > **Input integrity.** Run the checks in `references/data-input-integrity.md` before computing > anything, and report what they found. Each one produces a confident wrong answer rather than > a visible error, so a broken input does not announce it