Reservoir capacity is often sized from the historical inflow record alone, even though a single record is one realisation of an uncertain process. This study quantified storage–yield–reliability (S–Y–R) relationships for hypothetical water-supply reservoirs at six French gauging stations with contrasting regimes (CAMELS-FR data, monthly flows 1999–2018). Required storage was computed with the sequent peak algorithm (SPA) for draft ratios α = 0.3–0.9 of the mean annual flow (MAF) and for uniform (D1) and summer-peaking (D2) demands. Performance was evaluated by behaviour simulation using time-based reliability, resilience, vulnerability and a sustainability index. Three generators produced 1000 synthetic 20-year sequences per site: a log-space Thomas–Fiering (TF) model, an annual block bootstrap (BB) and an annual lognormal AR(1) model disaggregated by the method of fragments (AF). All generators reproduced the monthly means within 5%. Only AF preserved the lag-one autocorrelation of annual flows, which was high for the groundwater-fed Loing (r = 0.43). For α = 0.7 and D1, the historical SPA storage ranged from 0.27 to 0.62 MAF, and the storage for 99% time-based reliability ranged from 0.23 to 0.68 MAF (AF). At the Loing, AF required 81% more storage than TF and BB for 99% reliability at α = 0.7, showing that ignoring interannual persistence underestimates over-year storage. The summer-peaking demand increased the required storage by 12–34% at the five rain-fed sites but reduced it at the snowmelt-fed Ubaye. Reservoirs sized on the historical record achieved a mean reliability of 0.987–0.997 on the AF sequences, but the probability of no failure in a new 20-year period was only 0.42–0.81, with resilience of 0.22–0.46. The Gould–Dincer normal approximation deviated from the simulated median storage by factors of 0.54–11.9. Stochastic S–Y–R analysis with a generator that preserves annual persistence is recommended for sizing reservoirs from short records.