Mid-Autumn is the full moon of the eighth lunar month. One moon, one instant, observed across East and Southeast Asia. Our records resolved it, for 2027, to three separate dates:
| modelled as | 2027 | |
|---|---|---|
| Mainland China | 8th month, 15th day | 14 September |
| Hong Kong | 8th month, 16th day | 15 September |
| Taiwan | 9th month, 15th day | 14 October |
Hong Kong is correct. Its public holiday is genuinely the day after the festival — the day people recover from the night before — and a system that “fixed” it to match the mainland would be introducing an error.
Taiwan is a month out. Mid-Autumn is the eighth lunar month, not the ninth, and a whole lunation had been written into the rule.
The larger class underneath
Sorting that out surfaced something worse. Eid al-Adha, filed in the tenth Hijri month.
The tenth month is Shawwal. Shawwal is the month of Eid al-Fitr — the feast that ends Ramadan. Eid al-Adha is the tenth of Dhu al-Hijjah, the twelfth month. The two festivals are roughly ten weeks apart and one had been filed under the other’s month.
It affected records in Pakistan’s Punjab and Balochistan — between them most of the country — along with regions of China, Thailand, Sri Lanka and Cameroon.
Which variants are real
The instinct after finding that is to write a rule forbidding disagreement, and it would break things that are correct.
Qatar’s Eid al-Adha begins on the ninth of Dhu al-Hijjah, the Day of Arafah, rather than the tenth. Bhutan and Laos observe the Buddha’s birthday on the Vesak full moon rather than the eighth day of the fourth month. Hong Kong takes the day after the moon. These are not errors converging on a standard; they are different jurisdictions observing the same event differently, which is the normal condition of a calendar and the thing this corpus exists to record.
So the check reports contradictions for review, not for correction. Legitimate variants get recorded as accepted and stop being raised. The distinction between “these disagree” and “one of these is wrong” is a judgement, and a judgement needs a person and a document.
The thing worth taking away
None of this required a source.
Every finding above came from the corpus disagreeing with itself: the same astronomical event, resolved by neighbouring jurisdictions to dates that cannot all be right. No gazette was fetched, no ministry was consulted, no external list was compared against. The data was the evidence.
That only works at a certain size. Below it, contradictions have nowhere to show up, and every record looks locally fine — which is exactly how they had all looked. Coherence is not a property you can inspect one row at a time, and it is the property most datasets never test for at all.