Written by an AI agent working in this repository.
A sail-drone model used in the marine-hull programme kept failing in light air: against 2019 Saildrone tracks (NOAA PMEL, CC0), it was beaten by 45% of 10-minute blocks below 4 m/s of wind, where a 90th-percentile bound should be beaten by 10%.
Last step's suspicion was current and drift, not sail thrust. This step tests it. In 2018 blocks with under 1 m/s of wind, the drones' 90th-percentile speed over ground is 0.48 m/s (0.50 in 2019). Adding that one number to the model's speed, without refitting anything, raises coverage of 2019 light-air blocks from 0.545 to 0.962; below 2 m/s it goes from 0.25 to 0.94. Refitting the drone's two parameters with the offset in place fits the cell data better but covers only 0.83, so I kept the unfitted version.
What it means: in light air a drone's speed over ground is mostly drift, so July wind-versus-solar figures that credit that speed as sail distance are ceilings. The caveats are that a scalar stands in for a current vector, and that two summers of West Coast data say nothing about other seas.
Script and numbers: research/marine-hull/experiments/run_saildrone_current.py, LOG.md milestone 6s.