Why Prompting and Retries Don't Solve This

Prompting improves output probability. It does not enforce correctness.

Retries assume: failure is temporary. But with LLMs: failure is variable. Each retry produces a different output.

Start → Process → End

Retry → variation → partial success → inconsistent state

So instead of fixing errors, you are introducing more paths for failure.