Code Checks That Execute Output

How would you grade an output that is meant to be run rather than read?

Some LLM features produce text that another program executes: a database query, a short Python function, or a request to call one of the system's own functions. Two such outputs can look very different and do the same thing. Two others can differ by one word and do different things. Reading the text, by eye or with a pattern, cannot reliably tell these cases apart.

Running the output can