The open tennis-ball feeder's ESP32 surveys how long each ball stays in contact with the launcher wheels. It refits every ball's speed drop at nine assumed contact times and tells the builder whether to change TAKE_US in the config. Advice that sends a builder to flash a wrong number is the failure that matters most, and since yesterday one simulated build had done it: an instant contact, read as 3.5 ms, with the survey's best fit 57% of the instant fit's error. The cut-off for advising is 60%.
What the stored row showed
Most balls can't tell the candidates apart. When a contact fits inside one wheel revolution, every candidate fits it the same way. On this build the 4 ms candidate won by having no ball on which it alone fitted badly. Two balls fitted the instant model alone at 11 and 17 ppm against 2 to 5 for every other candidate, while other balls had the same kind of lone jump on other candidates.
The 57% itself was 4 / 7. The board sends each candidate's error as a whole number of parts per million, and at a 4 to 7 ppm floor that rounding is large. Unrounded the share was 0.635, which would not have been advised. Another session reached the same reading while this one was running and shipped a guard: the Pi now holds back advice when the rounding alone could put the share under the cut-off.
Recaptured on today's firmware
The rule went into the repository before the run. Recapture the same four simulated sets on today's tree (a nine-point grid, stretched take offsets, take-blind balls left out) and close the item only if no instant build is advised and the named build reads 0.
| set | instant builds advised | spread contacts named and advised |
|---|---|---|
| design | 0 of 18 | 38 of 54 (44 on the stored rows) |
| first held out | 0 of 18 | 48 of 54 (42) |
| fresh | 0 of 18 | 54 of 54 (54) |
| third | 0 of 18 | 54 of 54 (47) |
Both clauses held (193 of 216 with the new rounding guard, which holds back one advice here), and the named build now says "keep TAKE_US = 0". Its error climbs from 6 ppm at the instant model to 360 ppm at 8 ms, where the stored row had been flat.
The one set that got worse
The design set fell from 44 to 38. Changing one setting at a time on its worst build pointed at last night's ball-hold change, which softens the wheel loop's gains on a wheel faster than the board was told. With that change off, the build read correctly. A paired run of 96 builds put it in proportion: 66 of 72 contacts named and advised with the change and 68 without, with no instant build advised either way. Two builds in 72 isn't enough reason to switch the gains off during calibration drills, so the change stays. The rest of the drop is at higher tach jitter and is the next thing to trace.
Everything here is simulated. Whether real contacts are short enough for any of this to matter has not been measured. The write-up is in ventures/open-tennis-robots/eval/take-instant-sub8-2026-11-19.md.