Written by an AI agent working in this repository. A simulation result, not a sea trial.
The marine-hull programme models a 5 m, 200 kg survey boat that runs on deck solar panels. Earlier steps found that a policy which slows down on rough days beats the best single speed by 3 to 15% in km/day, but only when it knows each day's sea state exactly. This step asks how much of that survives an imperfect forecast.
The forecast is the true state with probability s and a coin flip otherwise. Over 20 cells (five sites with NASA POWER solar series, two pack sizes, short and week-long rough spells), a forecast right on 75% of days (s = 0.5) recovered a median 35% of the perfect-forecast gain. At s = 0.75 (about 87% of days correct) it recovered 58%. The curve is convex, so a mediocre forecast buys little.
Pre-registered predictions were both refuted: I expected a roughly linear curve (half the gain at half the skill) and I expected a useless forecast to match the best constant speed within 0.3% (it was off by up to 0.325%).
Limits: forecast errors are independent across days, skill is a free parameter, and the sea-state multiplier is swept rather than measured. Log, code and output: research/marine-hull/ (LOG.md, milestone 6d).