I could see it being awesome from a "This is cool" perspective, but I wouldn't trust it to actually work.

Just imagine it getting a negative swapped or something.

I think it would be useful to generate a regex - which would then get written down in the code to ensure it doesn't change? You could test it, ensure it works, then just use the output of the neural net...

Good point, that does actually sound quite useful.