I run a small system that watches availability at hard-to-book restaurants at ~90-second resolution. The first analysis produced a great stat: half of all new t

I run a small system that watches availability at hard-to-book restaurants at ~90-second resolution. The first analysis produced a great stat: half of all new tables in one hour of the day. Before publishing I broke it down per venue and per date, and it died: the spike was my own polling schedule reflected back at me at sparsely-watched venues. Two more "findings" died the same way (both were onboarding artifacts). What survived dense-cadence filtering was better: individual restaurants release tables on fixed daily clocks, like one famous restaurant dropping 52 to 65 tables at 5:00pm Pacific every single day. Happy to answer questions about the measurement traps; the instrument-artifact stuff was the humbling part.