Forecast-skill sensitivity ========================== The forecast-aware weather policy can be tested against a coarse forecast-skill model. With probability equal to the plotted skill, the annual forecast equals the actual calendar; otherwise it is an independent synthetic draw. Skill 0 is independent, skill 1 is perfect. This is a diagnostic for forecast error, not a meteorological forecast product. Results ------- 20 episodes, seed 7, 75-day fall window. .. list-table:: :header-rows: 1 * - Forecast skill - Completed scope P50 - Retained scope P50 - Mean risk - Cost P50 - Unscheduled h * - 0% - 86.2% - 88.3% - 0.0526 - 264,470 - 64.7 * - 50% - 88.0% - 88.8% - 0.0497 - 264,723 - 41.9 * - 100% - 87.6% - 87.6% - 0.0502 - 264,047 - 0.0 Unscheduled work falls monotonically as forecast skill rises. Completion and risk do **not** improve monotonically: mid-skill planning retains more scope and happens to complete more work, while the perfect-forecast case is conservative because it defers whole actions until the observed calendar fits. Limits ------ - The skill model is a perfect/independent mixture, not a calibrated forecast. - One annual forecast realization per episode; no within-year updates. - Whole-action deferral; no partial treatment credit. - No safety, asset-owner priority, access, or field-calibration claim.