The model is straightforward and, for most of the world, correct. A subdivision inherits the calendar of its parent and may add days of its own. Bavaria keeps the German holidays plus a few more; a Brazilian municipality keeps the federal list plus its own saint.
What the model cannot express is a subdivision that inherits a day and does not keep it — or keeps it on a different date. There is no way to say not this one, no way to say ours replaces that. Inheritance runs one direction only.
That is a modelling decision that was never quite made. It was assumed, because the additive case is overwhelmingly the common one, and it held until three places hit it in a single afternoon.
The three
Alderney takes its summer bank holiday on the first Monday of August. The rest of the Bailiwick of Guernsey takes the last. Alderney does not observe two summer bank holidays; it observes one, on a different day.
Sark keeps Liberation Day on 10 May. Guernsey’s is the 9th. Sark does not have two Liberation Days.
Malaysia is the same shape written into statute. The Holidays Act 1951 gives Chinese New Year “one day in the States of Kelantan and Terengganu, two days in the other States”, and Hari Raya Haji the reverse. The national record carries the majority case, and the two states have no way to express the minority one — which is not an exception the law tolerates but the arrangement the law prescribes.
A fourth, São Tomé and Príncipe, shows the boundary from the other side. Príncipe keeps three days the national calendar does not, and that works perfectly, because it is the additive direction. The same file still cannot say which national days Príncipe does not keep.
What we did instead of fixing it
There were two tidy options and we took neither.
Dropping the conflicting day would have produced a clean record that was wrong — a reader would see one date and no sign that another exists. Asserting the subdivision’s date as if the model supported exclusion would have produced a clean record that lied about its own structure.
So both days are present, and every affected record carries the conflict named in its own note: here is the inherited day, here is the local one, this is the operative date, and the reason you can see both is that the model cannot yet express what is actually true.
That is uglier. It is also the only version where a reader ends up knowing what we know.
Why say it out loud
A contract gap that gets worked around three times stops being a gap and starts being a habit, and the workaround is always cheaper than the fix in the moment it is chosen. The third occurrence is roughly where the arithmetic reverses, and writing it down is how you notice you have reached it.
The wider point is about the shape of inheritance itself. Most systems that model places as a hierarchy — tax rules, address formats, business hours, regulatory regimes — allow a child to extend a parent and quietly assume it will never need to refuse. Jurisdictions do not have that decency. A territory belonging to a larger one is not obliged to be a superset of it, and quite often is not.