Uncertainty and recourse
The prototype distinguishes several questions that are easy to blur together.
Factors and scenarios
Factors are multiplicative annual changes in cost, hours, or capacity. A scenario is one path through those factors. A band reports percentiles across a set of paths. A band is not a probability distribution unless the sampling design says so.
Static, robust, and chance plans
A static plan is fixed before the factors are observed.
A robust plan is selected against a set of sampled scenarios.
A chance plan limits violations on a sampled training set; out-of-sample violations can still occur.
A two-stage or recourse plan defers some actions after factors are observed.
A rolling policy re-solves future years as time advances.
The multi-year and weather pages compare these policies on common paths. They are decision experiments, not proofs that a policy is optimal or distributionally safe.
What recourse means here
Recourse is fixed-plan annual deferral or rolling re-optimization with a two-year lookahead. It is not a stochastic dynamic program. It does not compute an optimal policy over the full uncertainty distribution.
Weather, outage, and forecast skill
The weather experiments use an assumed calendar: five fixed outage days plus independent 20% daily weather loss in a 75-day window. The calendar-aware case observes the calendar before deferral. The forecast-aware case plans against an independent synthetic forecast and executes under the actual calendar. The forecast-skill sweep is a diagnostic, not a weather forecast product.
The weather model is documented in Weather/outage-aware recourse, Forecast-skill sensitivity, and Fall-window capacity remedy.
Claims to avoid
Do not report any of these as a result:
a distributional guarantee from a sampled chance constraint;
a globally optimal policy from a heuristic;
a field-validated safety or clearance standard;
a forecast service from the synthetic calendar draw.