What weather routing actually buys

A geometry / seasonal-β / forecast-α decomposition of optimizer fuel-ROI across ocean regimes — current-aware (ERA5 + GLORYS), calibrated MR tanker, 12 monthly departures (2025), hindcast. Confirmatory vs. a pre-registered protocol; nothing deployed.

Read the baseline and the foresight. Savings are reported against a constraint-respecting geodesic (shortest legal path), not the ship's planned route, so route-shortening is not miscredited as weather skill. Every number is hindcast (the optimiser sees the true weather) — an upper bound on operational, forecast-limited value; the tactical term α especially.

The decomposition

The total routing gain (planned route → spot-optimal), flown through the same realized weather + currents, splits additively into three components with very different robustness and commercial honesty:

$$ \underbrace{F_{\text{plan}}-F_{\text{spot}}}_{\text{total gain}} = \underbrace{(F_{\text{plan}}-F_{\text{geo}})}_{G\ \text{geometry}} + \underbrace{(F_{\text{geo}}-F_{\text{prior}})}_{\beta\ \text{seasonal+current}} + \underbrace{(F_{\text{prior}}-F_{\text{spot}})}_{\alpha\ \text{forecast (hindcast UB)}} $$

Routes

Honesty controls

Negative control (trade-wind, no dominant current): the method must show ≈0 weather gain — it does (weather term ≤ 0, no manufactured saving).  In-sample β: the season-prior is the medoid of this year's optima ⇒ β slightly optimistic, α slightly conservative; the clean estimator uses a multi-year climatology (specced).  Operational α: the realisable forecast-error number is lower than the hindcast α here and needs an archived forecast aligned to historical dates (specced, not run — forecasts are not fabricated).  Currents via CMEMS GLORYS; an earlier wind/wave-only run understated value on the current routes — currents are quasi-stationary (they enter β), so including them strengthens the structural>tactical conclusion.