Orange orbs are ordinary hazard events; magenta orbs are common-cause events that pass every layer. Grey = stopped. Hole size on a plate = independent failure probability q = (1-beta)p.
The science
Each of n layers fails (lets a hazard through) with probability p. In the beta-factor model a fraction beta of those failures is due to a shared cause (same training gap, same power supply, same pressure to skip checks): with probability c = beta p all layers fail together, otherwise each fails independently with q = (1-beta) p. The breach probability is exactly P = c + (1 - c) q^n. For beta = 0 this is p^n (each extra layer multiplies the risk by p); for beta > 0 it flattens at the floor beta p. Expected incidents per year = lambda x exposure x P, and P(at least one) = 1 - e^(-expected) (Poisson).
Honest limits. This is a textbook-style reliability toy: constant per-event probabilities, one shared cause, identical layers, no human learning or drift and no consequence model. Real layers have different p, partial dependence and time-varying failure; hazard rates, exposure and p here are illustrative, and the percentages are not benchmarks for any workplace. Training is modelled only as a lower p. It does not replace a real risk assessment.
Checks. A separate headless Monte Carlo (own xorshift RNG, 400 000 events, no animation) against the formula with a 4-sigma band, the independence limit, and the large-n saturation. The animated tally is shown with a 95 % Wilson interval. Scene: Blender-baked rig (rubber floor, brushed steel, plaster plates, wooden crate) in Three.js.