Optimal directions for searching without derivatives: an Evidence Press briefing. Some optimisation methods cannot use derivatives. Instead they try a small, fixed set of directions from the current point, and move if one of them helps. That only works if, whatever the downhill direction is, one of the trial directions points roughly the same way. The cosine measure scores a set of directions by its worst case: the cosine of the largest angle between any direction in space and the nearest trial direction. Bigger is better. The same number answers a question in geometry: how small can equal caps be, centred on those directions, and still cover the whole sphere? Two geometers, Böröczky and Wintsche, conjectured which few caps cover the sphere best. Split the dimensions into nearly equal groups, put a regular simplex, such as a triangle or a tetrahedron, in each group, and set the groups at right angles to each other. Hare, Jarry-Bolduc and Planiden asked a related question about positive bases built from simplices in complementary subspaces: must the best ones sit at right angles? This candidate proves that, within that class, they must, in every dimension and for every size. The cosine measure is at most one over the square root of the sum of the squared group dimensions, and only perpendicular regular simplices reach that bound. Arrangements that come close to the best are nearly perpendicular and nearly regular, at an exact rate. Going further, the perpendicular arrangement cannot be improved by any small change of the directions, even a change that leaves the structured class. So the conjectured best coverings are at least locally best, in every dimension. There is a twist. In four dimensions with six directions, a tetrahedron beside a perpendicular pair is also a local optimum, although two perpendicular triangles do better, so a local search can get stuck there. The proof picks a random corner of a related polytope and compares the farthest corner with the average. The paper credits this device to public figures by Jarry-Bolduc and Krishan Lal, who drew it for triangles and pairs. The global question for most sizes stays open, and the local result does not say how large the safe neighbourhood is. 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.