Fatal errors exist because some mistakes are not worth trading off. If an agent gives incorrect financial information, it does not matter that the greeting was warm and the call was efficient: the call was a failure and the score should say so.
The alternative, letting a weighted average absorb a serious failure into a respectable-looking 82%, is how QA scores stop being useful to anyone.
In QXAI
A parameter marked Fatal that scores below half forces the overall score to zero. This is applied on the server after the model returns, not left to the model to remember.