If it's a script, then u feed it to the interpretator. The interpretator processes all script commands one by one. If at some point it faces a problem (like division by zero), then it simply stops the processing.
Okay - so it *is* being interpreted (and presumably by each node that sees the new block also otherwise it could have been interpreted incorrectly).
I think you'll need a hash of the memory "state" after execution which should probably actually be provided by the person wanting to "execute" the script (presumably through a second tx).
To my way of thinking it would be:
1) Create the "script tx" and pay a fee just based upon its size (am not sure where we are storing its state - perhaps that is going to be in some sort of AM).
2) Send a "run script" tx later with a fee and an "expected state hash" (if the nodes run the script and fail to get the same state hash then it would be an invalid tx).
Or are you thinking along the lines of something more "automated" than this approach?
