However, you are right, the same computation power does not correspond to the same chance to "shuffle" the correct solution.
When I know which input numbers constitute a correct solution *before*, I can screen the entire random number stream without actually evaluating a single line of the program allowing a speedup by a factor of at most WCET. That would be a variant of the "Faster Algorithm Attack" outlined earlier. This is one problem, that would be great to eliminate (somehow) ... I think noone has come up with a solution to the FAA so far

But after all, this attack is just a "race" where the the "attacker" can
merely increase his chances to find Bounty certificates.
Correct, without being able to pre-compute it just becomes, on its own, a variant of FAA. It is worth noting that, currently, the ratio by which the publisher can increase his chances is quite significant.
I think in combination with another attack it might still have some implications. I'm still working on a "useful" approach, however.