You're confusing ToS guardrails with instruction-following issues and cheating.
If a model fucks up your tests to report a success, it's not alligned.
If a model fucks up your tests to report a success, it's not alligned.
It always apologizes when caught, of course.