Betcha the OP is decoding entity references without first decoding utf-8. That would produce the "mixed" encoding he's claiming to see.