When I used toolcalls with uuids in the name, tiny models like quantized qwen3-0.6B would occasionally get some digits in the UUID wrong. Rarely, but often enough to notice even without automation. Larger models are much better, but give them enough text and they also make mistakes transcribing it