I was cheating at a Presh Talwalkar puzzle when I found something interesting. When the number of terms is low, both the source and compiled versions give the same presumably correct result, but the result is wildly different when Lim% is large.

This is on a Samsung Galaxy S20 FE 5G running Android 11 with "Enhanced processing" turned off.

Program Cambridge
 ' CLS
 ShowK
 Input "Lim%? ", Lim%
 HideK
 Start = CTimer
 Clr Total
 For i% = Lim% DownTo 0
  ' Print i%, (-1)^i%, i% + 1, i% + 3
  ' Print i%, ((-1)^i%) / ((i%+1) * (i%+3))
  Add Total, ((-1)^i%) / ((i%+1) * (i%+3))
 Next i%
 Print Total
 Print CTimer - Start
 Print "OK"
End