Understand the range.
Plan for what matters.

Where the spread comes from, year by year

Change from each chain’s own 1990–2019 calendar-year mean. Dashed line: ensemble-mean smoothed response. Empirical bands are ensemble percentiles and the primary display. Normal-approximation bands are the mean ± 1.645σ from the variance split and assume a bell-shaped spread, which scenario mixtures and skewed runoff may not follow; dotted P05–P95 limits are drawn when the two disagree. Percentile curves are calculated year by year, so the chain forming P90 can change between years: they are not coherent drought trajectories. Internal variability comes from polynomial residuals.

P10–P90 cumulative response envelopes, fixed order. Labels show main-effect variance shares; increments depend on stage order. Interaction is reported separately. Percentages are shares of attributed variance; the bracketed range is the lowest to highest share across six alternative ways of separating forced change from natural variability, so none of them is a settled value (hover the headline for GCM-bootstrap intervals). Hydrological parameters are parameter sets within one model structure; model structure, groundwater and management assumptions are not sampled, so a small share here does not mean hydrological modelling adds little uncertainty overall.

A −20% change in annual runoff is a screening level, not an ecological limit, supply failure point or loss of resilience; those need a defined asset, performance measure and operating context. Shares describe this ensemble, not the probability of a scenario. Annual denominators count chain-years; other rules count chains. Each cell also states the number of chains.

Illustrative classes, not a real GCM selection. Counts describe model agreement, not scenario likelihood.