The science
risk(a) = 1 - exp(-lambda * a) leak probability of a copy of age a (days)
p = min(1, r0 / sqrt(R)) chance that an overdue rotation succeeds today
L = I - 1 + K, K ~ Geometric(p) days between rotations, E[L] = I - 1 + 1/p
mean risk = [E L - (1 - E e^(-lambda L)) / (1 - e^(-lambda))] / E L
overdue share = ((1-p)/p) / E L age distribution P(a) = P(L > a) / E L
lease risk = 1 - (1 - e^(-lambda T)) / (lambda T) age uniform on [0, T]
exposure = (R * raw risk + V * lease risk) / N
A hard-coded copy ages the moment it is created and only gets refreshed when someone finds every place it is used. Past the target interval each copy tries to rotate every day, but the success chance falls as the number of raw copies R grows (a coordination assumption, not a law of nature). A vaulted secret is a lease that expires after a few hours, so its age never grows and its risk stays near zero, while it also removes a raw copy and makes the remaining rotations more reliable.
Simplifications. Risk per copy is an assumed saturating exponential, not measured breach data. The rotation-success law p = r₀/√R is a modelling choice with r₀ as a free knob. Copies are independent; real leaks cluster (one leaked laptop exposes many secrets). Leases are assumed to be issued and renewed without failure; a real vault is itself a single point of failure that needs high availability and access control.
Independent numerical checks test the closed forms against Monte Carlo runs, direct summation, numerical integration and a full-system simulation.
Not run yet.