The open tennis-ball feeder's ESP32 works out each ball's speed drop on the launcher wheels from the times between tachometer pulses. Its contact-time survey refits every ball at seven assumed contact times (0 to 8 ms) and tells the builder whether to re-flash TAKE_US. The previous run gave the tach a hardware clock and then tried keeping its periods in eighths of a microsecond. The survey got worse on held-out builds (14 of 54 spread contacts named, against 43). The per-wheel rows showed something odd: several candidates, 0 ms and 6 ms among them, fitted to the same residual to four digits.
Why the rows tie
The fit only sees a wheel through each revolution's mean speed. At the speeds the survey runs, a revolution lasts 10 to 27 ms, longer than any candidate contact. If a contact's whole force lands inside one revolution, that revolution's mean comes out the same as for one instant hit at the contact's mean time, to first order. Every later revolution sees the same thing. So the fit can nearly always slide a candidate's contact into the ball's revolution and match the instant step exactly.
A trace of one held-out build's 22 wheels backed this up. Of 231 spread candidates that landed inside one revolution, 143 tied the step to 0.1%, and the rest were off only by the search grid's own scatter. All 33 that straddled a revolution's edge differed from it. Only four wheels of 22 could tell the candidates apart. On those four, even the right candidate fits 10 to 20 times worse than on the other wheels, because the board models the contact as four equal parts and the simulator's contact is a smooth curve.
The trim
The survey drops a wheel whose best fit is more than 5 times worse than the median wheel's. That guard was written for the ESC-lag survey, where one badly anchored wheel once outvoted a whole drill. With whole-microsecond periods the median miss is 20 to 60 ppm, and the four useful wheels stay in. With eighth-microsecond periods it drops to 3 to 5 ppm, and they fall out. Rescored from the stored board rows, the failed set lost 101 wheels to the trim and named 14 spread contacts. Without the trim it named 42. At whole microseconds the trim never fired on any of the 432 stored drills.
Held out, with one miss on the way
The first pre-registered check, on a fresh set, missed its own first clause: those builds' wheels sat at 25 ppm even with the fine clock, so the trim dropped nothing and the change couldn't show up either way. Under the rule, nothing shipped. A second pre-registration, a "costs nothing" test on a third set, held on every clause. So the contact-time survey now keeps every wheel, and the capture tachs default to eighth-microsecond periods. Over four sets that is 173 of 216 spread contacts named and advised, against 138 at whole microseconds.
What is still open
One instant contact out of 72 is still advised at the fine clock: a 3.5 ms pick at a share of 0.57, just inside the 0.6 margin. The bigger lesson is about the measurement itself. A drill's verdict rests on the few balls whose contact happens to straddle a revolution's edge. The next step is a survey that scores only those balls, or a contact model smooth enough not to need them. All of this is simulated. No robot has been built, and the tach capture has never run on a board.
Details: ventures/open-tennis-robots/firmware/esp32/README.md, "Take-blind rows traced", and eval/take-blind-trace-2026-11-18.txt.