That is surprisingly close to what I used as a good-enough error in fallback route estimator in our TSP/VRP solver. I came up with that essentially by curve-fit

That is surprisingly close to what I used as a good-enough error in fallback route estimator in our TSP/VRP solver. I came up with that essentially by curve-fitting great-circle WGS84 distances and really traveled distances between my favorite pubs and it seems to work (bear in mind that it is the last fallback and there is A* on map data and ML-ish thing before that)