The marine-hull research programme asks how much a wing-sail drone gains by planning its tacks from a wind forecast instead of following the wind shifts as they come. The previous step found that a forecast with error time constant of 30 minutes or less loses to the simple header-following rule, and that the planner tacked 150 to 250 times a day.
The obvious fix is a switch margin: tack only if the planned gain beats staying by some fraction of a step's progress. We pre-registered that it would halve the tacks at zero forecast skill and bring the planner level with the header-follower. It did neither. With margins from 0.02 to 0.2, tacks per day stayed at 166 to 173 (January) and fell only from 248 to 232 to 244 (July), and progress stayed 1 to 2.6 points below the header-follower. At useful skill (120 minutes) margin 0 was already best; margins of 0.1 or more cost 0.6 to 1.1 points.
So the chatter is not near-ties. Below about 30 minutes of forecast skill, use the header-follower. Data: 6-minute NOAA CO-OPS harbour winds, 48 windows per month, synthetic forecast-error model (not an archived forecast). Untested: a minimum-dwell rule. Log, code and numbers: research/marine-hull/LOG.md, milestone 5f.