\fp_eval:n {  exp(1) - (1+1/1009)^{1009}  } -> 0.001346
\fp_eval:n {  exp(1) - (1+1/1010)^{1010}  } -> 0.001613

\fp_eval:n {  exp(1) - (1+1/1011)^{1011}  } -> 0.001343


(1+1/n)^n should approach e monotonically, so there is clearly an error 
when n = 1010.


Similarly, there are errors when n = 1020 and 1032, and at 505 (half of 
1010, but 510 and 516 seem to be OK).


Andrew


PS. I haven't been testing l3fp integer by integer (truly). This just 
popped out at me when testing a LyX package based on l3fp that I've been 
working on.