The Chinese calendar was being computed by taking the tabulated date of Chinese New Year and adding, for each subsequent month, 29.53 days — the average length of a lunar month.
It is a reasonable-looking shortcut. Lunar months really do average 29.53 days. And it carries no concept of a leap month, which is the part that ruins it.
2025 has a leap sixth month. From the seventh month onward, every date computed this way landed about thirty days early. Mid-Autumn 2025 resolved to 6 September, while the State Council’s own holiday notice placed it inside the 1–8 October block. 2028, with a leap fifth month, failed the same way. And in ordinary years the truncation quietly shed a day.
One hundred and twenty-five of three hundred dates were wrong, across twenty-five jurisdictions. Thirty-nine of them by twenty days or more. Guangxi’s Zhuang New Year was landing in the wrong Gregorian year.
What the rule actually is
There is no shortcut, and the real rule is not complicated so much as uncompromising:
A month begins on the day of the new moon, reckoned in China Standard Time. The month containing the December solstice is month eleven. Where thirteen months fall between one month eleven and the next, the leap month is the first of them that contains no zhongqi — no major solar term.
Every clause of that is load-bearing. The calendar is not derived from an average; it is derived from where the moon and the sun actually are, resolved against a specific meridian, and the leap month is not inserted on a cycle but identified by a property.
The nine minutes
Computing new moons demanded more precision than expected. The abridged series we were carrying is accurate to within a few hours, which sounds ample for deciding what day something falls on.
It is not. In 2026, one month begins nine minutes from midnight, Beijing time. A calculation accurate to hours has no opinion worth having about which side of midnight that lands on — and which side it lands on shifts the whole month, and every holiday hanging off it, by a day.
This is the same lesson as the equinox, arriving from the other direction. There, a correct instant was dated in the wrong place. Here, an instant known only to the nearest few hours cannot be dated at all. A moment becomes a date only when you know both where you are standing and precisely enough when it happened.
Checking the fix against things that were not ours
The corrected computation was tested against Hong Kong’s gazetted 2026 list — seven of seven — and against known Mid-Autumn and Dragon Boat dates from 2025 to 2028.
Then against an existing seventy-one-year table of Chinese New Year dates: it matched seventy. The disagreement was 2053, where the table gives 19 January — a date outside the range Chinese New Year can occupy at all. The computation was right and the reference was wrong.
Which is worth stating plainly, because the instinct on a 70/71 result is to assume the single mismatch is yours. Sometimes it is the table. The way to tell is to know what the answer is constrained to look like, and Chinese New Year cannot fall on 19 January any more than Christmas can fall in June.