The open tennis-ball feeder's ESP32 measures each ball by how much its launcher wheels slow down as the ball goes through. The previous run found that fitting 24 revolutions of the wheel's recovery, instead of 8, cut the error of that reading by a quarter on a jittery tachometer. The obvious follow-up was the board's other fit on the same data, the contact-time survey. It refits every wheel at seven assumed ball contact times (0 to 8 ms) and tells the builder whether to re-flash TAKE_US.
Two pulls
A longer window gives the fit more rows to average the tach's noise against. But a ball's contact time only changes the first few revolutions after it. Every extra row is one where all seven candidates fit alike, which pulls their scores together. The board's advice only speaks when the best candidate beats "instant" by a clear margin (a best/step share of 0.6 or less). So diluted scores mean silence.
What the simulator said
tests/take_survey_window.py keeps every survey wheel's inputs exactly as the board passed them. It refits each one at 8, 12, 16 and 24 revolutions, scoring them with the board's own sums and margin. The builds: four contact times on three plants, one of them a 3 µs jittery tach that no take-survey run had tried, and three launchers. The design run (36 drills) showed the dilution clearly, so I wrote down a prediction and a decision rule and committed them before running a held-out set at a new wheel lag and seed:
| window | contact times named and advised (of 27) | instant takes wrongly advised |
|---|---|---|
| 8 (shipped) | 12 | 0 |
| 24 | 9 | 0 |
Every clause of the prediction held. On the 50 ms ESC plant the mean share rose from 0.47 to 0.63. Its 3 ms builds were advised 2 times in 3 at 8 and never at 24, though the picks themselves stayed right. So the survey keeps 8, now a named constant (TAKE_SURVEY_AFTER), and the drop reading keeps 24. The two fits ask different questions: one gains from averaging, the other depends on the rows right after the ball.
The finding that matters more
On the jittery tach the survey was blind at every window. Shares ran from 0.90 to 1.00, and picks scattered: an instant take read anywhere from 0 to 5.5 ms. The margin kept it quiet each time, which is the right behaviour. But it means that on a real wheel the survey will only ever say something if the tach is well under 3 µs of edge jitter. The next step is to sweep the jitter down to find where the survey starts to see, and then log a real Hall sensor on a spinning wheel.
Everything here is simulated with one seed per set and an assumed contact profile. The numbers are in ventures/open-tennis-robots/eval/take-survey-window-2026-11-18.txt, and the firmware README's "The take survey's window" has the full tables.