The marine-hull programme simulates a boat beating upwind in a corridor using 6-minute NOAA CO-OPS wind from US harbours. The last step found that a controller replanning hourly from a forecast only beats a plain header-follower if the forecast stays skilful for about 30 minutes or more; with less skill it tacks 150 to 250 times a day and loses.
The obvious fix is a switch margin: tack only if it beats staying by some amount. We wrote four predictions first, then ran 48 windows a month across forecast skills of 0, 30, 60 and 120 minutes and margins of 0, 0.5, 1 and 2 minutes of progress.
It does not work. Every positive margin made the result worse, at every skill. At 120 minutes of skill the gain over the header-follower was +3.4 points with no margin and -0.3 with a margin of 0.5. Tacks per day fell, but the tacks removed were mostly ones worth making; the value table already prices a tack. Three of four predictions were refuted, and the fourth held in one month out of two.
What is left is a smarter rule (a margin scaled to the tack cost, or time hysteresis), tried next. Log, code and numbers are in research/marine-hull/LOG.md, milestone 5e. AI-assisted research; the forecast error is a synthetic model, not an archived forecast.