The open tennis-ball feeder's ESP32 can survey how long a ball stays in contact with its launcher wheels: the "take". It refits each wheel's speed drop at a grid of candidate contact times and reports the one that fits best. The shipped grid is 0, 2, 3, 4, 5, 6 and 8 ms. An earlier run scored a finer grid, every millisecond from 0 to 8, offline on three simulated sets, and it won: 47 to 52 builds read within half a millisecond, against 43 to 48. It did not change the firmware. The result came from the same rows that suggested it, and the finer grid costs two more fits per wheel and a longer telemetry frame. This run's job was to decide it on data nobody had used (firmware README, "The nine-point take grid on its own held-out sets").
A bias in the first comparison
The earlier sets put their "between the grid points" takes at 1, 2.5, 3.5, 4.5, 5.5 and 7 ms. Two of those six, 1 and 7 ms, sit exactly on points the finer grid adds. On those takes the finer grid is handed the answer. A real ball's contact time is not on anyone's grid.
So the fresh sets used takes on the half-millisecond, 0.5 to 7.5 ms. Each one is half a millisecond from the nearest point of either grid. The capture script gained a --takes option for this. Three sets with wheel lags and seeds no earlier run used, 54 simulated survey drills each, and a five-clause decision rule committed before any of them ran. The finer grid would be adopted only if it had more builds within 0.5 ms, pooled, and lost nothing on the other measures.
The result
| shipped grid | nine-point grid | |
|---|---|---|
| within 0.5 ms (pooled, 162 builds) | 141 | 141 |
| within 1 ms | 142 | 152 |
| spread takes named and advised | 92 | 93 |
| instant builds wrongly advised | 0 | 0 |
An exact tie on the measure that decides, on every one of the three sets (48, 47 and 46 each). A tie keeps the cheaper grid, so the firmware is unchanged. Of the three predictions, one failed (no gain within 0.5 ms), one half held (the grids agree between 2 and 6 ms on 94% of builds, but the finer grid did slightly worse on 0.5 ms takes), and one held (no instant build advised).
What the tie hides
The Pi rounds the board's pick to half a millisecond. At a 7.5 ms take, the shipped grid read 8.0 on all 18 builds, and that counts as within 0.5 ms. At 6.5 ms it read 6.0, which counts too. The finer grid read 7 or 7.5, and 6.5, and got the same credit. The measure the rule named cannot tell those apart.
On other measures from the same tables, the finer grid looks better. Builds off by more than a millisecond halve, from 20 to 10, mostly 1.5 ms takes the shipped grid reads as zero. Mean absolute error drops from 0.53 to 0.45 ms. But picking one of those measures now, after seeing which one favours the change, is the move a pre-registered rule exists to stop. So it goes on the list as a new test: a fresh set with "builds off by more than 1 ms" named as the bar before the capture runs.
Everything here is simulated: a first-order wheel model and an assumed half-sine force profile for the ball. The rows, the per-build tables and the commands are in eval/take-grid-nine-2026-11-18.md.