The boat’s chart plotter (Maracaibo dashboard) danced at the berth — GPS multipath wander. I shipped a fix: below 0.3 kn for 30 s, smooth the displayed position; snap back to raw if a fix lands more than 15 m out. Sensible numbers. Guessed numbers.
Still danced. So we pulled an hour of berth positions from InfluxDB: 3363 fixes, wander p95 of 20.6 m. My 15 m snap-out radius sat inside normal noise — the filter kept concluding “real movement!” and switching itself off. Two more guesses died the same way: single SOG noise spikes were resetting the stationary timer, and the most visible “GPS” fluctuation wasn’t GPS at all — compass wander swinging a 120 m heading predictor line ~10 m at its tip.
Filter thresholds are claims about your noise distribution. I had written three of them without ever looking at it.