When the starting line matters: fixation in noisy swarm decisions. An Evidence Press briefing. First, the background. Robot swarms often choose between options, such as two nest sites, without a leader. In the weighted voter model, each robot advertises its favourite for a while, then copies a random neighbour. Better options are advertised for longer, so they spread. In 2025, Ganesh, Hauert and Valla conjectured a formula for the chance that a small group backing the better option eventually wins in a large swarm. Their formula describes the other robots as they look after a long run, in equilibrium. Yet the conjecture says every robot starts with a freshly measured quality. Now, the main finding. That difference matters, and the conjecture fails as stated. A snapshot of a long-running swarm over-represents slow robots, because slow robots keep their state for longer. This is the inspection paradox. A fresh swarm has no such bias, so its robots switch faster at first, just when a small group is won or lost. So whenever advertising rates vary, the better option wins more often than predicted. In one example, the true chance is about seventy-eight per cent, not seventy-six. The source's own simulations started in equilibrium. There, the formula is exactly right when the better option has one fixed rate, but it underestimates again when that rate varies. The worst-case gap approaches twenty-seven percentage points, and a short burn-in removes it. The limitations: the proofs cover complete networks, and only when every better-option rate is below every other rate. This is an unrefereed, AI-assisted candidate, released on twenty-seven October twenty twenty-six. The paper, proofs and checks are linked on this page. This briefing uses a synthetic AI voice, and is not additional evidence. This briefing uses an AI-generated voice and is not additional scientific evidence.