Don't Give the Model SQL

My health data has six traps in it that have each already produced a wrong answer. Given SQL, a model walks into all six. Told about them in a prompt, it avoids them most of the time, which is worse.