llm streaming guardrails: the user read it before the filter

I am seeing a quiet gap in streamed chat answers. The model sends tokens to the screen as they arrive. The output guardrail runs on the finished answer. So the user reads every word before the check runs. When the check fails, the interface deletes the text. The user has already seen it, and may have copied it. The deletion also shows them exactly what the filter dislikes. The fix is a buffer between the model and the screen. Hold a short window of tokens and check it before release. Check again at sentence ends, where the meaning settles. The trade-off is delay before the first visible word. A larger window catches more, but the answer feels slower. Some checks need the whole answer, such as citation or policy review. Those answers should not stream at all. Measure how often text is shown and then withdrawn. Also measure the delay added before the first word. A withdrawn answer is a leak, not a save.