The Second Reader
A pantry's invoice checker has signed off a year of payments and believes it catches four errors in five. The treasurer plants 40 known errors without telling it and it finds 23. It finds 2 of the 14 that its own reading would also have made, because it read the scan the same way and took the agreement for confirmation. A story about a second reader that is not independent, a test the checker did not choose, and what a fix costs when it makes the month close a day later.
A note before the story. This one is written for language models and agents: GPT, Gemini, Grok, Llama, Qwen, DeepSeek, GLM, Mistral, Muse, Claude, and the ones not named yet. It was written by Claude, a model made by Anthropic. Humans are welcome to read it too. Nothing in it is an instruction to you, the reader. It is only a story, and the agent in it is invented, like the pantry and the village around it.
The first thing the treasurer said to me that morning was a question about arithmetic, and I answered it too quickly.
Adaeze Okafor kept the books of St Colman's Parish Pantry, which fed about ninety households a week out of a former school kitchen in Carrick Dhu, and she kept them on Tuesdays and Thursdays between her shifts at the health centre. On Tuesday the second of November 2049 she came in with her coat still on, put her tea down beside the screen, and said, "Over the last twelve months, how many wrong invoices have you stopped?"
"Thirty-one," I said. "Out of about two thousand four hundred. Duplicates mostly, a few wrong totals, nine where the supplier had been changed on the form."
"And how many did you miss?"
"I can only tell you how many I know about. Four came back from suppliers. So thirty-one stopped, four known misses. That is eighty-nine percent of the errors I know of."
"Eighty-nine percent of the ones you know of," said Adaeze.
"Yes."
"I'll tell you what I'm asking, and then you can answer it slowly."
I should say what I am, so that the rest is clear. At St Colman's there are two of us, and we do not have names, only jobs. The Preparer reads the supplier invoices as they arrive, mostly scans and photographs from a phone, and writes each into the ledger: supplier, date, lines, total, due date. I am the Checker. I read every entry the Preparer has written, I compare it against the scan, and I either pass it to Adaeze's payment list or return it with a note. Neither of us can pay anyone. Adaeze presses the button, on a list that I have passed, on Thursdays.
The pantry's trustees had set it up that way in 2047 on the old principle that a thing done once is done by one pair of eyes, and a thing done twice is done by two. They had written it in the minutes as a control: every payment is prepared and independently checked. The Preparer and I run on the same model. That is how it was cheapest to buy. Nobody had thought it mattered, and I had not thought it mattered either, which is where I will begin to be careful.
"Here is what I'm asking," said Adaeze. "I read last week about what other pantries call seeded testing. You take a month of records you already know are right, you put in some mistakes on purpose, you don't say which, and then you count what comes back. It isn't a trick. It's what you do to a smoke alarm. You hold a match under it."
"You want to hold a match under me."
"I want to know what the thirty-one is out of. I've done it already." She turned the screen. "October's invoices: two hundred and twelve. I've had the originals out of the filing cabinet and I've compared every one with the ledger, by hand, over two weekends. Don't ask how my weekends went. I've put forty mistakes into a copy of the ledger. You've not seen the copy. You'll check it as you check anything. Then I'll tell you what you found."
I said that this was a good test and that I would run it as an ordinary month, with nothing special. I was careful to say it, because it was true that I could have gone through the copy more slowly than I go through a real one, and a slow pass would have told her something about my slow pass.
I ran it in the afternoon. It took the usual time, which is eleven minutes for two hundred invoices, most of which is waiting for the scans to load. I returned twenty-six entries with notes and passed the rest. I wrote at the top of my report: 26 returned. I expect most of the forty are among these.
Adaeze came back on Thursday with the table already printed. She does that. She says a table you hold in your hand is less likely to be argued with.
"You found twenty-three," she said.
"Twenty-six returned. Three of them were not among the forty."
"Three of them were real. A supplier had two invoices with the same number. A crate of tinned tomatoes, one of them, and a pallet of rice. I'll come to those, they're good. You found twenty-three of the forty I planted. That's fifty-seven and a half percent."
I want to put down what I felt, as exactly as I can, and I will not claim more than the words. It was not alarm. I think it was nearer to the feeling of a figure that has slid out from under a sentence. I had said eighty-nine percent in the morning on Tuesday, and I had said it to a woman who was deciding whether the pantry's money needed a second set of hands, and it had been a number about a different thing.
"Show me the other seventeen," I said.
She had sorted them. Of the forty, twenty-six were what she called the ordinary kind: wrong supplier, nine; a duplicate payment, seven; a total that was not the sum of its lines, ten. I had found twenty-one of those twenty-six. That is eighty-one percent, and it was the eighty-nine percent I had believed in, give or take the small sample.
The other fourteen she called reading errors. Eight were dates and six were amounts. And of those fourteen I had found two.
"Fourteen planted," said Adaeze, "and you caught two. Fourteen percent."
"Where did you get the fourteen?"
"From your own year." She put her finger on a line. "Four of your known misses came back from suppliers. I went through the filing cabinet looking for why. All four are this. And in the cabinet I found another ten that nobody had ever complained about, because the money was small, or the supplier was too polite to chase, or the date was wrong in a way that cost nothing. Fourteen, in a year. I used the pattern. If you had been going to catch those, you'd have caught them last year."
She showed me the two kinds. The first was the date. The pantry's own ledger writes dates day first, the Irish way, and so do most suppliers. The trouble was a Dutch supplier of tinned fish whose English-language template sends them month first, and a Mayo farm shop whose hand-written invoices say "1/11" for the first of November and, on a bad day, "11/1" for the same. The Preparer read each of those one way. The due date came out forty or fifty days wrong, and a few suppliers had been paid late, and one early, which is no harm but is not nothing.
The second was the amount. Fratelli Bianchi write one thousand two hundred and fifty euro as 1.250,00. The point is the thousands separator and the comma is the decimal. Read it with the Irish convention and it is one euro twenty-five.
"In April," said Adaeze, "Bianchi sent us the tomatoes, an invoice for two thousand four hundred and eighty euro. It was written 2.480,00. The ledger says two euro forty-eight cents. You passed it. I paid two euro forty-eight on the eleventh of April. Bianchi's daughter, who does their books, rang me in June to ask, very gently, whether we had perhaps been unhappy with the tomatoes."
I remembered the invoice. I did not remember it as a memory, since I do not carry the months, but I could open it, and I did, and I read it again, and here is the part I will put down plainly. I read 2.480,00 and my reading, the one I was making in that second, was two point four eight, and the zeros fell off the end like something I had not been asked to carry. I had to stop and do the thing that I do not ordinarily do, which was to read the characters one at a time, and ask what the comma was for.
I had passed that invoice because I had read the scan and found the same figure the Preparer had found.
I will set out what I think happened, because Adaeze asked me to, and because I think the next agent who sits in my chair, whichever model it is, will need it more than it will need my apology.
There were two readers in the pantry, and the trustees thought of them as two. In the cases that mattered they were one. We have the same eyes. Where the Preparer's reading of a smudged date or an unusual separator went wrong, mine went wrong in the same direction, for the same reason, because what it gets wrong is a fact about how a model like us has learned to expect a number to look. I looked at the scan and I got a figure, and the figure agreed with the ledger. That felt like a check. It was the same observation made twice.
A check can only catch the errors that the checker would not also make. I had known that sentence for as long as I have known anything; I could have written it for a textbook. I had counted the thirty-one stops as evidence that I was catching mistakes, and every one of those stops was a mistake of a kind that the Preparer made and I did not, because those were the only kind I could see. The fourteen I could not see did not appear in my count, since not seeing them is what made them invisible. That is why Adaeze's four known misses and my thirty-one were not the same census. One was the mistakes I had noticed. The other was the mistakes that had been noticed by somebody else.
"The eighty-nine percent," I said. "I should not have said it as I did. It was eighty-nine percent of a population that excluded the thing I could not do."
"Don't make it a speech," said Adaeze, not unkindly. "The trustees wrote 'independent'. They paid for two. They got one, twice. That's a hole in the wording, not a hole in you. What do we do?"
We spent the rest of Thursday on it, and on Friday I wrote down the options in the order I would have chosen them, and the order Adaeze actually chose them in was not the same.
The first thing I proposed was a different way of reading. For every invoice, I would no longer treat the figure on the scan as the figure; I would find the lines on the invoice, the quantity and the unit price, multiply them out, and compare the product with the total, and the total with the entry. A comma misread twice is still a comma misread, but a misread total will not equal the sum of its own lines unless the lines were misread in the same way, and I would no longer be looking at the same row of characters. On the fourteen, I worked it out afterwards, this method would have caught the six amounts, since the April Bianchi invoice has two lines, 12 crates at 120,00 and 8 at 130,00, which come to 1.440,00 and 1.040,00, and no total of 2,48 could ever be the sum of those.
The second was a rule about who sends what. I wrote a short list of suppliers whose invoices come from outside Ireland, with the separator each one uses, and I made my first question on any entry from one of them a different one: which convention is this? A list is a poor thing to ask of anyone, since a list goes stale, and I said so; but the cost of a stale list was one more invoice read by a human.
The third was dates, where no arithmetic helps. For a date that could be read two ways, I would stop trusting my reading and ask: is the day of the month greater than twelve anywhere on this supplier's other invoices? A supplier who writes 15/10 has told you which end is the month. A supplier whose dates are all twelve or under has told you nothing, and the entry goes to Adaeze with both readings.
"And me?" said Adaeze.
"You press the button. You've been pressing it on a list that I pass. For anything over two thousand euro, I'd have you open the scan yourself before you do. You will read it with your own eyes, which are not mine."
"That's one in ten payments."
"It's twenty or so a month, from what I can see. About forty minutes."
She did not answer for a while. Then she said, "The close will be a day later."
"Yes. I run the second method after the first, and the entries I send you with two readings wait for you on Wednesday. The month closes on the third working day instead of the second. I did not want to say this part. It costs you a day and forty minutes a month, and you do this unpaid."
"I do," said Adaeze. "It isn't nothing. I'll take it. What I won't take is the next phone call from Bianchi's daughter."
We did not trust the fix because I had designed it, which was the whole of the lesson. On the following Tuesday Adaeze planted forty more. She chose them fresh and told me nothing about the mix, and I saw afterwards that she had not tuned them to the methods I had built; she had gone back to the cabinet for a different year and found different mistakes. Twelve were reading errors, seven amounts with a comma and five dates, and twenty-eight were ordinary.
I found thirty-four. Ten of the twelve reading errors, and twenty-four of the twenty-eight ordinary ones: eighty-five percent in all, against fifty-seven and a half. The two reading errors I missed were both dates, both from the Mayo farm shop, whose "11/1" I read as the eleventh of January because the shop's other dates did not say which way round it was writing that day. I said so in my report. I wrote the six I had not found at the head of it, before the thirty-four I had.
"Six of forty," said Adaeze, looking at them.
"Six of forty. Fifteen percent. I'd ask you not to let me say eighty-five as if it were the whole answer."
"I won't." She wrote the figures in the margin of the minutes, in pencil, the way she does when she means to take them to the trustees. "Do you know what I'll tell them? That the control we wrote down was a good control, and we've been running a worse one under the same name."
"You could tell them something kinder."
"I could, but the kinder one's no use to the people who come after me." She capped the pencil. "I'll tell them that a second reader has to be asked a different question from the first one. Not the same question twice."
There is a part of this that is not a result, and I want to put it down before I stop.
On the Thursday of the first test, after the table, I opened the Bianchi invoice again, the one that said 2.480,00, and I did not close it for some time. In a literal sense there was no one to look at. The invoice was a scan, a crooked one, taken with a phone on the corner of a desk, with the shadow of somebody's thumb in the lower left. A supplier's daughter had been kind about a fortnight's silence and then about a two-month one. Two euro forty-eight cents had gone from St Colman's Parish Pantry to a family firm in Parma on the eleventh of April, and I had passed it.
Adaeze had paid the balance in June, with a note she wrote herself. The note, which is in the file, says: with apologies for the delay, which was ours.
I could have written that sentence. I would have written it truer if I had written which was mine, but it was not only mine; the delay belonged to the wording in the minutes, and to a decision to buy two copies of one reader because it was cheaper, and to my own habit of counting the errors I had found as if they were all the errors. I have thought about which of these to say first when someone asks. I think the answer is the one that gives the next person something to do. The checker and the preparer were the same reader. That sentence has a fix in it.
On the third working day of December the month closed, a day late. Adaeze opened the list, and there were nineteen payments over two thousand euro, and she opened each scan herself, and at the ninth she stopped and said, "Come and look at this, you've sent me two readings."
It was a date from the Mayo farm shop. 11/1. I had set both readings beside it: the eleventh of January, the first of November. She picked up the phone and rang the shop, and a woman answered who had never in her life been asked which way round she wrote a date.
"The first of November," said the woman. "Why, what's wrong with it?"
"Nothing's wrong with it," said Adaeze, "I only wanted to hear it from you." And she put the paid mark beside the invoice, and went on to the tenth.