I've been thinking about how I would rig the machines IF I had the ability. I wouldn't add a fixed number of votes onto a candidate because, at the lowest level, its too easy for the votes to outnumber the electors, as mentioned for "one Ohio precinct" above.
My first inclination would be either to count every vote for the person I didn't want to win at, say, 0.95 of a vote and the desired winner at 1 for each vote, but any audit investigating the software (should it come to it) would immediately declare using non-integer counts of the number of votes suspect. More subtle would be to count each nondesired vote as 95 and each desired vote at 100 and divide by 100 somewhere. However, as the machines use a mechanical tachometer, I understand, this would also be difficult.

No, the easiest and safest way is to either omit a proportion of the nondesired vote or, better, transfer it to the desired vote (this is better because it would mean the total number of votes cast equals the number of voters, allowing another audit test to be passed.)

All this is, I suppose, fairly obvious. But what you could then do is ask what proportion would have to be transferred to make one state (shall we say Florida?) give the same result for the reported vote as the exit poll. You could then try exactly that same proportion of transfer in all the other states and see how well they then match the exit poll. If they were much closer than the 'real' results, the evidence of something going on would be extremely strong.