The open tennis-ball feeder's ESP32 estimates how long each ball stays in contact with the launcher wheels. To do that it refits each wheel's speed drop at seven assumed contact times (0 to 8 ms), adds up each candidate's misfit over the wheels, and tells the builder whether to re-flash TAKE_US. It gives that advice only when the best candidate fits clearly better than the instant step: its mean misfit has to be at most 0.6 of the step's.
The last run found why that verdict was so often silent. A revolution of the wheel lasts 10 to 27 ms, longer than any contact. When a ball's whole contact lands inside one revolution, every candidate fits exactly as well as the step. Only the few balls whose contact straddles a revolution's edge can tell the candidates apart.
Wheels that cannot move the pick
A wheel that fits every candidate alike adds the same amount to every candidate's sum. It can't change which candidate wins. What it does is drag the best-to-step ratio toward 1, and that is the number the advice is gated on. A drill with twenty blind wheels and three informative ones reads as "no clear winner" even when the three agree.
The board can't cheaply work out from the tach edges which balls straddle, but each wheel's own row already shows it. So the take survey now leaves out a wheel when its worst candidate's squared misfit is within 1.5 times its best. It needs no extra memory and costs one min and one max per wheel. A wheel it leaves out still counts toward the Pi's wait. If every wheel so far is blind, the Pi says so and tells the builder to keep TAKE_US. It doesn't guess.
Chosen on old data, tested on new
The value 1.5 was picked by rescoring the board's own rows from all ten stored capture sets (tests/take_survey_blind.py). Every one of those sets had been seen before, so their numbers count as design: spread contacts named and advised went from 330 to 398, and the one wrongly advised instant build went away. Then the rule was written down and committed, and two fresh sets were simulated at a lag and seed never run before:
| fresh set, 72 drills | named and advised (of 54) | instant builds advised |
|---|---|---|
| hardware-timed capture, 1/8 us | 54, then 54 | 0, then 0 |
| jittery tach (the board's default) | 12, then 26 | 0, then 0 |
Every clause held, so the filter ships on. The hardware-capture set was already at its ceiling and could only show the filter doing no harm. The gain is on the tach the board actually runs today, and it comes at 0.5 to 1 us of edge jitter. From 1.5 us of jitter upward the survey is still nearly blind, naming 1 of 9.
What this does not show
All of this comes from the simulator's first-order wheel, one seed per set, with an assumed contact profile. A real wheel's blind rows won't tie as exactly, so 1.5 may need retuning on a board. A verdict now also rests on fewer wheels: the jittery set left out 1,149 of 1,584. The 16-wheel minimum before advising still counts all of them, and making it count only the kept wheels is the next item. Details are in the firmware README, "The take survey skips take-blind wheels", and in eval/take-survey-blind-2026-11-18.md.