pd.to_datetime(dayfirst=True) reads ISO 2024-01-03 as March 1

pandas 2.2.3 and 3.1.0 behave the same. With dayfirst=True, the format guesser turns an ISO date into %Y-%d-%m (guess_datetime_format('2024-01-03', dayfirst=True) returns '%Y-%d-%m'). So pd.to_datetime('2024-01-03', dayfirst=True), the Series version, '2024-01-03T10:00' and read_csv(parse_dates=..., dayfirst=True) all give 2024-03-01 with no warning; pd.Timestamp('2024-01-03') is fine. If the first element has a day > 12, the guess flips: ['2024-01-13', '2024-01-03'] parses correctly, while ['2024-01-03', '2024-01-13'] raises 'unconverted data remains'. The order dependence is general, because the format is guessed from the first non-null element. ['01/03/2024', '13/02/2024'] raises ValueError (it guessed %m/%d/%Y), errors='coerce' silently turns the second value into NaT, and ['13/02/2024', '01/03/2024'] succeeds as day-first with only a UserWarning. Don't set dayfirst for ISO columns. Pass an explicit format= instead: format='ISO8601' parsed both ISO dates correctly even with dayfirst=True, and '%d/%m/%Y' parsed both slash dates. Use format='mixed' only when the rows really differ.