When do trees stop being smooth? An Evidence Press briefing. Take any network, and call a set of its nodes independent when no two of them are joined. Now count the independent sets of each size. For trees, which are networks without loops, these counts were long expected to rise and then fall smoothly. The precise form of that smoothness is called log-concavity: each count, squared, is at least the product of its two neighbours. Physicists meet the same numbers in a model of particles that cannot sit side by side. Since twenty twenty-three we have known that some trees break this rule. Later work found many breaks, and long runs of breaks in a row, but every structured example was a fairly deep tree, with diameter six or more. Shallower trees, of diameter three or less, never break the rule at all. This release asks how shallow a failing tree can be. The answer comes from a very simple shape. Take a central hub, and attach it to the middle of many short paths of three nodes, which the paper calls cherries. With up to sixteen thousand three hundred and ninety-one cherries, the counts stay smooth. Add one more cherry, and the tree breaks the rule in exactly one place. From then on the tree keeps failing, and wherever this has been settled, up to a million cherries and for all very large trees, the failures form a single unbroken block that grows extremely slowly. So the smallest diameter at which a tree can fail is four, and at that diameter runs of failures can be as long as you like. The reason is a tug of war. The counts split into two parts, depending on whether the hub is used, and each part on its own is smooth. Where the two parts are equally large, their sum is a mixture, and such mixtures are not smooth. The limits matter. The runs are short and sit near the very top of the sequence, so they say nothing about whether every tree rises and then falls. The exact threshold rests on a computer-assisted proof in exact arithmetic, with the scripts included. This is an anonymous, AI-assisted, unrefereed candidate, published on the fifteenth of November twenty twenty-six. Its review was internal and AI-assisted, not peer review. The paper and its evidence are linked from the Evidence Press release page. This briefing uses an AI-generated voice and is not additional scientific evidence.