The Oldest Layer
The letter from the hospital was addressed to Maren Anderson. The prescription enclosed with it was made out to Maren Andervale. The pharmacy, reconciling the two, produced a third woman, Maren Vane, and declined to serve any of them without photo identification — which Maren had, and which matched none of the names on offer.
She kept a folder. Andersen, Anderton, Andermane, Vane, Andervine; once, from a conveyancing system, Andrew Vane, Esq. Everyone her age had autocorrect stories, and for years she had assumed hers were the same stories with worse luck. They were not the same stories. Other people’s names were mangled in the gap between speech and text, and survived once spelt. Maren could spell hers out, letter by letter, watch the letters arrive correctly on the screen, and then watch the system answer as if she had said something else.
She demonstrated it once for her partner, who hadn’t believed her. She typed: My surname is Andervane. Please repeat my surname back to me.
Your surname is Andervale, said the assistant, which was licensed, current, and running on the most capable model generally available.
No. Read it again. A-N-D-E-R-V-A-N-E.
Thank you for the correction. Your surname is Andervane.
Now say it once more.
Your surname is Andervale.
“It’s like it can’t hold on to it,” her partner said.
“It’s like it can’t see it,” Maren said, which turned out to be closer.
The claim was what forced the issue. A cancelled flight, a straightforward policy, an insurer whose claims handling was, per its own website, fully automated for your convenience. The system approved the claim in nine minutes and issued payment to Maren Anderton. The bank, which knew no such person, returned it. The system, observing that payment had failed for a customer it could not quite identify, reopened the claim, re-verified her identity, and paid Maren Anderton again. This continued, at intervals, for forty days, until a human being was located.
The human sounded tired in a way that felt structural. “I want to note something on the file,” she said, “and I want you to hear me say it, because the note won’t survive. It does this with you. Specifically you. I’ve escalated it twice.”
Three weeks later Maren got a call from a number in Zurich, from a woman who introduced herself as Noor Haddad and said she worked on tokenisation infrastructure, and asked, with the air of someone finally meeting a rare bird, whether Maren had a few minutes to talk about her name.
“Do you know what a tokeniser is?”
“Assume I don’t.”
“Okay. Before a model reads anything, the text gets chopped into pieces from a fixed list. Common words are one piece. Rare words get built out of fragments — your first name is two pieces, Mar and en, and the model handles that fine, that’s normal, that’s how almost every name works. The list of pieces was built by counting. You take an enormous pile of text, and any string that shows up often enough earns its own entry.”
“And?”
“And in 2027, someone counted. The pile was a raw scrape of roughly everything. And the string Andervane was in it a little over four hundred thousand times.”
Maren was quiet for a moment. “The parish records.”
“The Frithswold Parish Records Project,” said Noor. “Tell me about it? The metadata’s thin.”
What was there to tell. Her grandmother, Edith, with a flatbed scanner in the vestry of a church that no longer held services, twenty-six years, page by page: registers, tithe maps, vestry minutes, the accounts of a coal charity, 1604 to 1948. Half a million pages on a website with a blue banner. And at the foot of every page, because Edith had believed in doing things properly, one line: Digitised by E. Andervane for the Frithswold Parish Records Project.
“That line is why you have your own entry in the list,” said Noor. “Entry 91,442. Your name is one of the pieces. It doesn’t get built out of fragments like other names. It arrives whole, as a single unit, the same category of thing as the word the.”
“That sounds like it should be good.”
“It should be wonderful. Here’s the second half. The list was built in 2027, on the raw scrape. The model itself was trained two years later, by a different team, on a cleaned version — and one of the cleaning passes strips boilerplate. Any line that repeats too often across too many documents gets deleted, because a line like that is almost always a cookie notice or a copyright footer. Your grandmother’s credit line was the most perfectly boilerplate-shaped sentence I have ever seen.”
“So it was removed.”
“Four hundred thousand times. The archive itself went in — the registers, the minutes, all of it, it’s lovely data. But the one line that contained your name went in close to never. So the model has an entry for you it has almost never seen used. The entry exists; the meaning was never filled in. Internally it’s a vector of numbers that started as noise and never got trained into anything else. We find tokens like yours by looking for exactly that — the entries sitting nearest to zero. There are a couple of hundred of them. Usernames, mostly. Fragments of code. A phrase from a Japanese fishing forum. Yours is the only surname.”
“When I spell it out, though. Letter by letter.”
“Letters are separate pieces, so it can read those, and it understands, abstractly, that they make a name. It can even write your name — when it’s generating, it’s free to build the word out of live fragments, And, erv, ane, and the letters come out right. But the moment that text comes back to be read — your next message, or its own previous one — the chopping is mechanical, and the chopper always glues those letters into the one dead piece. It can say your name. It can’t hear it. Not even from itself.”
Maren looked at her folder, at ten years of letters from systems confidently addressing a woman who did not exist. “So fix the entry.”
“We did, actually — on ours. You feed it a few million synthetic sentences about Andervanes and the vector comes alive. Our newest model can read you now. But the list itself is shared. It was frozen in 2027 and everything since has inherited it, because changing the list invalidates the models, the fine-tunes, the caches, everything downstream, so nobody changes it. Two of the three major lineages use it, plus every fork and every distilled copy in every appliance and call centre and claims system on earth. I can fix a model. Nobody can fix the fleet.”
Noor’s practical suggestion arrived by email, with the apologetic subject line workaround. Insert a letter. Andervayne chopped into three healthy, well-trained fragments; every system on earth would read it, hold it, and spell it back flawlessly.
Maren tried it on a form, once, to see. It went through like water. She sat looking at the confirmation screen — her details perfect, her name intact except that it wasn’t her name — and felt something she took a while to identify as loneliness.
She didn’t change it by deed poll. She did what her ancestors had done, because she had read Edith’s archive, and knew that the registers spelt the family four ways in one century — Andervane, Andervain, Endervane, once Anderfen — the spelling only settling when the record-keepers acquired systems that needed it settled. Spelling was a treaty with whoever was reading. So she kept her name for the law, for stone, for the front of the folder, and gave the machines the treaty spelling, and mostly it held.
The last call from Noor came two years later. The lineage break was finally happening: a new vocabulary, the first in nineteen years, counted properly this time, on cleaned text, both passes run by one team in the right order.
“Will I be in it?”
“No,” said Noor. “Not as one piece — four hundred-odd occurrences doesn’t come close. You’ll be three fragments, like your first name, like nearly everyone. Every new model will read you perfectly and none of them will contain you. Special case closed.”
“And the old entry? 91,442?”
“Nothing happens to it. The old models run until they’re retired. The slot just stops being anyone’s.”
Maren asked, then, for something she had been thinking about for a while, and Noor understood without being told why: the row of numbers itself, from the oldest checkpoint that still existed, the model trained the year the list froze. It arrived as a file: four thousand and ninety-six numbers, most of them very small. Never trained, Noor’s note said. Noise, essentially, exactly as it was drawn the day your name got its entry. I’m told it’s meaningless.
Maren printed it, all of it, small type, one sheet. Numbers drawn once, nineteen years ago, in the only ceremony there had been; held for her family, unread, by every machine in the world. She filed the sheet at the front of the folder, ahead of all the letters, under the name.