It comes to my mind that if recycle is not turned off during the tests, it could explain the variance. So I did the test, ans yes, it's a lot better. Here the new version:
comp: func [n [int!] a [block!] b [block!] /local ref][ print reform bind [product version platform build] rebol recycle/off ref: perf n  a: to-decimal probe abs (perf n probe a) - ref b: to-decimal probe abs (perf n probe b) - ref recycle/on to-percent (to-integer b - a / a * 100) / 100 ]
Again I fucked up ref calculation. Lot more stable now
perf: func [i [integer!] b [block!] /local t n] [ n: 0 t: stats/timer bind b 'n ; to use vars i,n in b loop i [do b ++ n] stats/timer - t ] comp: func [n [int!] a [block!] b [block!] /local adjust][ print reform bind [product version platform build] rebol recycle/off adjust: n / to-decimal perf n  a: adjust * to-decimal probe perf n probe a b: adjust * to-decimal probe perf n probe b recycle/on to-percent (to-integer b - a / a * 100) / 100 ]
make it run for at least a couple seconds each if you want a stable result.
Gregg: it seems your machine has several process that take CPU in the background. Hence the funny results. Has Gab say: make a longer loop (1M) to reduce this effect... or kill process ;-) Could be funny to 'comp the exact same code and has so different results.
Maybe Gregg's machine is reporting to NSA, FBI, CIA and other agencies in the background? :-)
the code i used on R2 was:
time*: func [n block /local start] [ start: now/precise loop n block difference now/precise start ] time: func [block /local count time result] [ time: 0:00 count: 1 while [time < 0:00:01] [ time: time* count block result: divide to decimal! time count ; multiply by two for faster convergence ; (ie. aim for 2 seconds) count: 0:00:01 * count * 2 / time ] result ]
>> time [1 + 1] == 7.25799625998013E-8 >> f: does [1 + 1] >> time [f] == 1.39604952270899E-7
How do I save the contents of dir-list to a file from the console ?