Fixed-size subset sums: an Evidence Press briefing. Pick some numbers from a list at random and add them up. How predictable can that sum be? This candidate answers exactly, in three senses. First, how often can the sum be zero? If the numbers add up to zero and are not all zero, and you pick k of the n numbers, the paper proves that at least two times n minus two choose k minus one of the picks give a nonzero sum, whenever k is at least five and n is larger than twice k. Only one pattern achieves this: a single plus one, a single minus one, and zeros everywhere else. In graph theory, this is the smallest footprint of a vibration mode of the Johnson graph. Mogilnykh, Vorob'ev and Valyuzhenich conjectured it in twenty eighteen, after checking every n up to six hundred by computer. Second, how predictable can the whole distribution be, measured by Rényi entropy? A single outlier gives the most predictable sum, except when exactly half the numbers are drawn: then two outliers win, once the order of the entropy passes an explicit threshold. So a natural Bernoulli lower bound holds for every population exactly when the order is at most two, and fails beyond it. Third, how rarely can the sum be nonnegative? Aydinian and Blinovsky conjectured that lists with just two distinct values are the worst case for the Manickam, Miklós and Singhi conjecture, which asks how many picks from a zero-sum list must have a nonnegative sum. For those lists the paper proves the expected bound whenever n is at least four times k, where earlier work needed eight times k, using a recent lemma of Ai, Karagiannis, Kuželka and Pelekis. So their conjecture would imply the full Manickam, Miklós and Singhi conjecture; this does not prove it. The limits matter: the zero-sum theorem is for real numbers, only the first eigenspace is treated, and the computer checks supplement the written proofs. Evidence Press published this anonymous, AI-assisted, unrefereed candidate, and its review was internal, not peer review. This briefing uses an AI-generated voice and is not additional scientific evidence.