It would be an act of faith to believe any of the last 2 ms digits, even on an incremental basis, without good evidence that they are unaffected by process scheduling. Anymore, when people get serious about measuring elapsed time accurately, they arrange to read the high-resolution timer hardware available in most CPUs. An extension function which read that hardware, (perhaps using \<chrono\> in C++), would be a better bet than trying to push SQLite's timestamp facility beyond what it is designed to do and do so across multiple platforms.