What I was looking for
When to stop collecting order-flow data, decided before looking at it.
What I found
"Keep collecting and see" never ends and invites stopping when the answer looks good.
In the order-flow study I set a target of ±0.05¢ and recomputed, at each checkpoint, how many more observations each estimate needed: 1.6× more for the 30-second markout, 10× more for the 300-second one. It showed early which estimate was the bottleneck.
The move
I set the confidence target first, then compute how much more data each estimate needs, so collecting has an end date.