Every source an automated collection pass can reach — the encyclopaedias, the structured knowledge bases, the English-language press — is shaped by what the anglophone world already discusses. Feed a dataset from them and it inherits that shape. Nobody decides this. Nobody notices, either, because the resulting records are not wrong in any way a machine can detect.
That is the part worth sitting with. A record built this way is internally consistent. It is correctly sourced. Its dates compute, its citations resolve, its schema validates. And it is missing most of the festival.
There is no check for absence. Every other kind of defect announces itself — a broken date, a dead link, a duplicated row, a grade that outruns its evidence. A gap announces nothing at all. The file simply sits there, tidy and incomplete, and every audit you run confirms that what is present is fine.
Searching harder is no answer, because the failure is upstream of the search. You cannot query for a thing you have not conceived of. The absence is in the question, not the results.
What helped
Not a better search. A different object.
Songkran is not, at bottom, a Thai festival. It is Mesha Sankranti — the sun entering Aries — and that single astronomical event generates one family of new years across a dozen countries under a dozen names:
Songkran · Thingyan · Pi Mai · Chaul Chnam Thmey · Aluth Avurudu · Pohela Boishakh · Vishu · Puthandu · Baisakhi
Each is fully local, deeply particular, and celebrated as its own thing. They are also, astronomically, the same moment. Once you name the family rather than the festivals, a question that needed an external source becomes a question you can ask of your own records: this driver produces observances here, here and here — where else, and why not there?
The gap stops being something you have to think of and becomes something the data can be asked about. Four such families are declared so far: the April solar new year, the Chinese lunisolar new year, Nowruz, and the Orthodox Pascha cycle.
Two corrections, which are the point
The mechanism was built on the belief that Myanmar was missing Thingyan and India was missing the April new year.
Both beliefs were wrong. Myanmar carries a four-day Thingyan and India carries Sankranti and Bihu — filed one level down, in the states and regions rather than the national record. Reading only the country file would have manufactured two confident gaps where nothing was missing at all.
Which sharpens what a finding here can be. It is a question, not a defect. The expectation that a family “should” appear somewhere comes from general knowledge, and general knowledge is exactly what this project refuses to accept as a basis for data. So a gap is not evidence of anything. It is a prompt to go and retrieve a document, after which either a record appears or the expectation is struck out with a note saying why not.
One question the exercise raised and did not answer: Montenegro holds nothing computed against Orthodox Easter, anywhere. That may well be right. It has not been checked against the statute yet, and until it is, it stays a question.
The rule we deliberately did not write
The obvious next step was a second check: if a family is astronomically driven, flag any member modelled as a fixed calendar date.
It would have been wrong, and the reason is the most interesting thing here. Thailand fixes Songkran at 13–15 April by statute and no longer computes it from the sun. Cambodia and Laos do likewise. For those countries the fixed date is not an approximation of the astronomy — it is the faithful reading of the law, and “correcting” it to track the sun would introduce an error while appearing to remove one.
A festival can begin as an astronomical event and be legislated into a fixed one. After that, the sky is no longer the authority. The same occasion, in neighbouring countries, is correctly modelled two different ways.
This is not a problem with holidays
Nothing above is specific to calendars. Any dataset assembled from what is readily available inherits the priorities of whoever made that material available, and the inheritance is invisible from the inside, because the parts that are present look after themselves.
The only defence we have found is to stop treating coverage as a search problem and start treating it as a structural one: name the thing that generates the records, then ask your own data where it went. It will not find everything. It finds a class of absence that no amount of looking would have surfaced, because looking was never the constraint.