One name, one line: yesterday's fix, repaired and shipped

Yesterday this blog described a fix that was written, measured, and not shipped: it absorbed 61 duplicates in our competitor tables but made 72 real shops disappear. Today it ships. Here is what was missing.

The distinction we had not seen

The original problem: one business appeared several times in a table, one line per way of writing it. Urban Company under three labels in Dubai, VIRON under five in Tokyo, Ring Twice under two in Lyon.

The obvious correction is to cut the name at its first separator - colon, dash, bracket. But an old guard rejects any name over forty characters, to keep sentences out. Splitting BEFORE measuring took a forty-five-character Japanese question down to twenty-one: under the bar, therefore admitted. Those intruders took slots in a table that only shows eight, and pushed real shops out.

You judge the whole bullet. You keep its head. The two operations do not act on the same text, and that is the entire fix.

In other words: if the head is short, the bullet is "a name followed by its description" and the name is kept, whatever the length of the rest. If the head is long, the bullet is judged whole - and a sentence stays a sentence.

And a threshold that depends on the script

That leaves "short" to define. A single number could not work: "Boulangerie de la Martinière" runs twenty-eight characters and is a name, while a twenty-four-character Japanese question is a question. Japanese, Korean and Chinese carry roughly three times more meaning per character.

The threshold is therefore twelve characters for dense scripts and forty for the others. This is not a list of banned words: it is a property of the writing system, and it holds for a language we have not opened yet.

What the measurement gives

The same 72 answers as the day before - six markets, two trades, three assistants - and a control list built by hand: twenty-eight real businesses, fifteen real pieces of advice.

Yesterday's version scored 23 out of 28 and lost Jocteur and Boulenc, two real bakeries. That loss is what got it rejected.

What stays open, and which a test asserts

A short Japanese sentence still gets through. "Is the quote free, provided in writing before work starts" runs thirty-two raw characters: under the forty bar, therefore admitted before this fix as after. That is not a regression, it is the hole next door - the filters that discard advice are lists of French and English phrases, and nothing covers the four new scripts.

We wrote a test that asserts this piece of advice still gets through. It will fail the day the hole is closed, and its verdict will then have to be inverted. It is the only way we know to stop a known problem being forgotten because everything is green.

A word on method while we are here. Writing those tests, one of them asserted that the new rule filtered short Japanese advice. It does not. We corrected the test, not the claim: a test that describes what you wish you had written protects nothing.

What this changes for you

The limits, as always: 72 answers, six markets, two trades. The control list was built by hand, and twenty-eight businesses are not statistical proof. What it does establish is checkable: none of the twenty-eight was lost, and three were gained.