greatest

of the nature of the job at hand. I am writing numerical/scientific code and the problem suffers from a very well understood and documented loss of numerical precision (it is about computing the motion of satellites and planets using Kepler's laws - and the numerical inaccuracy happens near escape velocity). My testing is perhaps closer to stress testing than unit testing. Here's my test framework (it's only about 300 LOC): https://github.com/rikusalminen/libkepler/blob/master/test/n... https://github.com/rikusalminen/libkepler/blob/master/test/n... Here's a few example test cases: https://github.com/rikusalminen/libkepler/blob/master/test/k... The method is 1) create a pseudo random initial setting 2) perform a calculation in two different ways and verify that they match 3) analyze the results and verify that laws of physics apply 4) repeat (about 8 million times by default). My tests use about 15 minutes of CPU time currently, which gives me reasonable confidence that everything works as intended.