Common pitfalls

Updated Aug 23, 2026
  • Loading two language skills in one turn. Pick the one whose decision surface dominates and keep the other in reserve.
  • Routing an obvious five-line invariant through a verification ceremony instead of loading hypothesis directly.
  • Replacing a finite normative parameter table with generated values. The table is the specification.
  • Treating mutation testing as a build gate. The survivor list is a worklist; CI should publish the trend, not fail the build.
  • Profiling without a baseline. Pyinstrument is most useful as a before-and-after comparison.