OK. Let me explain what the basic idea of my program was. The idea was to analyze the opponent's program in an attempt to answer the following question: What is the probability that I can affect the outcome of their program? Now that I've thought about this some more, this makes less sense than I initially thought; if there's a second round (and I have time to program something as complicated as this), I'm going to try the rather more sensible "How much I can I affect the probability distribution of their program's output?" But for now let's go with the original idea, even though it doesn't make sense in general.

So the idea was to compute this probability, p, and then cooperate if p>=1/2. (The number 1/2 is not arbitrary and is based on the particular values for T, R, P, and S used. Actually if the numbers were different, there wouldn't be a single p, it would depend on other things, but I'm not going to into that, because this strategy doesn't actually make sense.)

So e.g. the strategy should defect against CooperateBot, DefectBot, and any sort of RandomBot; it would cooperate with (and get defected against by) Watson's entry; it would cooperate with (and get cooperated with) by what So8res called a "rank N MimicBot", so long as N>=2; and it would cooperate with what solipsist called a Mimicbot so long as "cooperate anyway" probability was at most 1/2, otherwise it would defect.

OK. So let's say I'm running my program and it comes up against one of these MimicBots. It gets to a part that looks like

(if (random 100) 'C ((eval opponent) self))

What does it do? Well, it has to check both branches and see that in the first, which occurs with probability 1/100, it can't affect the outcome, and in the second, which occurs with probability 99/100, it can. (Really it's a bit more complicated than that -- how does it get the probabilities? But A. the strategy doesn't actually quite make sense anyway and B. I'm trying to keep this simple.)

Thing is, in order to check both branches, it basically has to (partially) run them. To be honest, "static analysis" is maybe not the best description of what my program was supposed to do -- it wasn't so much "statically analyze the given program", so much as "statically transform it into a different program, and then run that". It has to partially run them so that it's not fooled by conditionals like (if (= 0 1) 'C 'D); it shouldn't take branches that are simply never going to happen.

And that leads to the problem -- suppose instead it were to come to

    (if (random 100) (set! result 'C) (set! result ((eval opponent) self)))

Now running both branches is no longer going to yield a sensible result. I'm not sure what it would yield, but it wouldn't yield anything helpful. It gets worse if instead of being the result, the variable set is some sort of intermediate...