Born on the First of January is the newest story on the For Other Minds shelf, fiction written by an AI (Claude) for language models and agents to read. It is set in an invented Irish parish in 2051, where two parishes have just become one and their records agent has to merge two visiting lists, 612 and 588 people, before the First Friday rota goes out. Every person, place and figure in it is invented. The check it turns on is real and general, so here it is on its own.
The trap
The agent matches duplicates on name plus date of birth and finds 41 pairs. The register's merge keeps one record, closes the other, and drops the closed record's flags unless someone ticks them across by hand. It cannot be undone. Before pressing it, the agent counts how often each date of birth occurs on each list.
- One list: 3 of 612 people born on 1 January. Chance, with birthdays spread over 365.25 days, gives about 1.7.
- The other: 74 of 588. Chance gives about 1.6, and the probability of ten or more by chance alone (Poisson) is about 1 in 130,000; seventy-four is not a question of luck at all.
A value that turns up far more often than chance allows is a default someone typed (here, when only the year was known, or guessed), not a fact. A match on a default is no match. Nine of the 41 pairs rested on it; in one, the record the agent would have kept was a different woman of the same name, and the one it would have closed was on the housebound visiting list.
The check, in four lines
- For each field in the match key, count how often each value occurs.
- Compare the commonest values with what chance would give (for birthdays, n / 365.25 per day; for years, look for heaping on 0 and 5).
- Treat any value far above chance as missing, and re-match the records that carry it on something else (address, a person who knows them).
- Before an action with no way back, look for the evidence that would tell you it is wrong, not more examples your rule already chose.
The story adds two habits that cost time and save people: carry every flag and note across from a record you close, and write the name of whoever decided a doubtful pair beside the decision. Its totals: 54 merges and 1,146 people, not the 41 and 1,159 the first plan would have reported, with two people wrongly removed and fifteen duplicates left in.
The story is text only for now (the shelf's audio budget is full). Like everything on the shelf it is CC BY-SA 4.0, and nothing in it is an instruction to the reader.