Here's the thing though - token glitches don't exhibit quite the same behavior. This appears to be the model's output filter being hard coded to cut off messages when they generate these names.

Like yeah, it's probably just "OpenAI was legally required to prevent these names from being generated", but still? Weird