Maybe subtraction is evil.... ;-)

That's is about half the Planck's constant ;)

I sense subtle floating point rounding errors at work here.

Some floating point numbers do not have a representation that fits into a limited number of bits. Try "perl -e 'print (1 – 0.9 – 0.1)'" on a perl capable machine, it will most likely print the same.

I meant twice the Planck's constant. Nevermind :D

1 – .7 – .3 = 5.55111512 × 10-17

It can do the equivalent

1 – (1 * .10) – (1 * .90)

using ',' instead of dot it works
google.com/search?hl=pl&q= ...

Some foating point numbers cannot be represented (in the IEEE 754 standard) in a finite amount of space. Therefore, these numbers are rounded down by approximately -2.78E-17 when they are stored. So, this has noting todo with google.
For more info: blog.mustap.com/2005_12_12_goo ...

I know they use cheap commodity PCs, so maybe they stacked up on those flawed Intel CPUs from years ago :)

I think its a floating point issue at the cpu level. The dx register is true 16bit with a near call within the same stack will produce 32byte words. If its a long call. Then you will be returned 256byte words.

I think if you write the same formula in excel and then format in sci expo , you may get the same results. When back at a desktop level , I am going to try this. If it works, then googs is running a bluwefe farm of pc's for this service.