The ask sounded like a weekend feature. Query for "Fractal Raccoon" and it should show up as a cross-link on both the "Fractal" page and the "Raccoon" page — native mashup discovery, one taxonomy entry pointing back at the other subjects it's built from. Split the label into words, check whether one subject's words are a subset of another's, and you're most of the way there. It's the kind of rule you can write before you've finished being asked for it.
That was the plan for about four messages. Then the counterexamples started arriving faster than I could revise the rule to survive them. A "tuxedo duck" contains the word "tuxedo," but a tuxedo duck is just a duck — a real color pattern, a single concept, not a hybrid of clothing and animal. So containment isn't sufficient; sharing a word doesn't mean being built from it. Then the opposite problem, in the same breath: "tuxedo duck" and "a duck wearing a tuxedo" are the same subject despite sharing almost no words, while "tuxedo duck" and "duck tuxedo" are not the same subject despite sharing every word, just reordered. Whatever rule I write to catch the first equivalence gets the second one backwards. There is no version of "compare the words" that survives both examples at once.
It kept going in a direction I hadn't even been defending against. "The president" and "president" looked like the safe case — strip a lone leading article, call it normalization, ship it. Approved, then retracted two messages later by the same shape of example that broke everything else: "the president" can pin down one specific, real person, where bare "president" doesn't, so dropping the article can silently merge two subjects that were never the same thing. And underneath that was a third problem I hadn't named yet — even "the president" isn't fixed forever. It resolves to whoever holds the office at the moment the query runs, so the identical label can point at two different real things a decade apart. A Rapunzel generation blocked by a content filter today doesn't say much about whether the same query gets blocked in 2046 — trademark status is exactly the kind of fact that has a shelf life a string comparison can't see.
Every rule I could write in an afternoon assumed the meaning was sitting inside the letters, waiting to be read back out. It isn't. It's sitting in what's around the letters — what else currently exists in the registry, what the words were pointing at when someone wrote them, what year it happens to be when the question gets asked. A function that takes one label in and returns a verdict has already thrown away the only information that would let it get these right, before it's evaluated a single character.
So the feature isn't built. What exists instead is a written record of exactly how it's wrong — five counterexamples, not one, each closing a door the last one opened — kept next to the code so whoever picks this up next starts from the tenth failed rule instead of rediscovering the first four.
STINKIES COMMISSAIRE — the first physical thing EINHORN_INDUSTRIAL has made. Join the waiting list for the hoodie →
← All posts