Getting Real Work out of a Model

The Department of Computer Applications at a government polytechnic in Coimbatore runs a forty-hour employability module for final-year students. Placement season starts in November. For two years the module’s most popular handout has been four pages titled “50 Power Prompts”, collected by a lecturer, Mr Balan, from social media posts and a video series. It opens with “You are a world-class career coach with 20 years of experience”, includes “take a deep breath and work through this step by step”, and ends with a note that offering the model a two hundred dollar tip improves results.

Six hundred copies had been quoted for by the campus press. The head of department, Dr Selvi, had approved the spend. Printing was booked for Friday.

On the Tuesday a student called Aarthi came to Mr Balan with a problem he was not expecting. She had used prompt 14 (the world-class career coach one) to write a covering letter for a testing role at a firm in Chennai. The letter was fluent and said nothing. She had then written four lines of her own naming the advert’s three requirements and the one project she had actually done, pasted her CV underneath, and asked again. The second letter was better, and she wanted to know why the handout had not told her to do that.

Mr Balan did the thing the handout did not. He took ten past adverts and ten real CVs from students who agreed to it. He ran the handout’s prompt over all ten, three times each, and scored every letter out of four on a checklist: does it name a requirement from the advert, does it cite something from the CV, does it avoid inventing experience, is it under three hundred words. Then he ran a plain version with no persona, no breathing, no tip, with the advert and the CV pasted in full.

The handout version scored 21, 19 and 22 out of 40. The plain version scored 31, 33 and 30.

Neither total was the interesting part. The interesting part was that three runs of the same unchanged prompt differed by three marks on their own. That meant any single before-and-after demonstration (the kind that fills the posts the handout was assembled from) could show whatever the person running it was hoping to see, and could show it honestly, without anybody lying.

He found a second thing he had not gone looking for. Two students had drafted in Tamil and asked for a translation, and both had run into the tool’s message limit on a document that in English would have fitted easily. The same passage in Tamil was costing roughly three times as many tokens. The students working in their first language had a smaller working space than the students working in their second, and nothing in the interface said so.

So there was a decision, and both sides of it cost something real.

Reprinting meant rewriting the handout in three weeks, which the campus press could not absorb. The next slot was late November, by which point the module would be half over and six hundred students would have started placements with no material at all. Part of the approved spend had already gone on plates.

Not reprinting meant handing six hundred students a document that Mr Balan could now demonstrate was worse than a page of plain advice, in a module whose entire purpose was to get them hired.

Dr Selvi’s objection was not that he was wrong. It was that “we measured it and the magic words did nothing” is a hard sentence to say about a document the department has taught from for two years, and that the handout was popular precisely because it looked like secret knowledge. A page that says paste the advert, paste your CV, name the hard part and say what you will not invent looks like something the students already knew.

The legal-aid desk runs on Saturday mornings out of two rooms behind a law college in Pune. Six student volunteers see about thirty people a day, most of them tenants with a notice to quit, a withheld deposit, or a landlord who has stopped repairing something. A visiting advocate, Ms Kulkarni, supervises and signs off anything that goes out on paper.

The volunteers had started using a chat tool between visitors, and the answers were poor in a way that took Ms Kulkarni a month to name. A volunteer would type something like: the landlord is being unreasonable about the deposit and the agreement seems to say he can keep it, what are her rights. Back would come a competent paragraph about deposits in general, which the volunteer would then repeat to a woman who had brought her agreement with her, in a folder, and put on the table.

Every important word in that prompt was the volunteer’s own reading. Unreasonable. Seems to say. The agreement was two feet away and had not been shown to anybody.

The obvious fix was to paste the agreement. That was also the thing Ms Kulkarni could not simply authorise. These are real people’s names, addresses, employers and bank details, handed over in confidence at a desk that exists because they cannot afford a lawyer. Pasting them into a commercial chat product means the text is stored on somebody else’s systems, may be retained, and is outside anything the desk can promise a visitor.

Both sides of that had a cost that was easy to state and hard to weigh. Describing the document produced advice about a category rather than a case, and the desk had already, twice, told somebody the wrong thing about a clause nobody had read out. Pasting the document exposed a person who had trusted a room of students with their tenancy.

They settled on a protocol before they settled on a tool: names, addresses, phone numbers, employers and account numbers replaced with bracketed placeholders, the agreement pasted as text, and nothing pasted at all for anything touching a criminal matter or a child. Ms Kulkarni tested it on her own tenancy first and found that the structure of the document survived redaction completely, because the parts that carry the meaning are the clauses, not the names.

The protocol worked and immediately produced a different failure.

A volunteer uploaded a photographed agreement (eleven pages, taken on a phone, slightly skewed) and asked whether it said anything about arbitration. The answer was that it did not, and that the tenant was therefore free to approach the Rent Authority directly. Ms Kulkarni, reading the paper copy over the volunteer’s shoulder, found an arbitration clause on page nine.

Nothing had gone wrong with the reasoning. The product had split the upload into chunks, indexed them, and put in front of the model only the passages it judged relevant to the question. Page nine had not been retrieved. The answer was a confident, well-formed statement about a document the model had mostly not seen, and there was nothing in the interface to say so.

That is when the desk stopped treating an attachment as a document.

The second protocol was three lines long and had to be run before any substantive question: how many pages did you receive, what is the first line on page one, and what is the last line of the document. Then, for anything about a clause: list every section heading in order with its number, and give the total count. Only after that did anybody ask what the agreement actually said, and the answer was required to quote the whole sentence including any words like unless, except, provided that or save where, because Ms Kulkarni had by then collected four examples of a summary that reversed a clause by dropping its qualifier.

The cost of all this is nine or ten minutes per file, on a Saturday when thirty people are waiting and the queue is visible through the door.

The desk handles about four hundred WhatsApp messages a day from retailers across Gujarat and Maharashtra, and roughly two thousand a day in the three weeks before Diwali. Four staff sort them by hand into five buckets (new order, order status, return, payment, and everything else) and paste each into the right sheet. In the rush, messages sit unsorted for six hours and retailers ring the owner’s mobile.

His nephew Jignesh, who is doing a BCA, built a classifier over a fortnight. It is a prompt with five labelled example messages, a list of the five categories, and an instruction to output the label and nothing else. He tested it on twelve messages he wrote himself and got twelve right.

The operations head, Mrs Shah, had two objections and only one of them was about the model.

The first was that four of the five examples were returns. She had not read a paper about class balance; she noticed it the way anyone would, by reading the prompt and finding that it looked like a prompt about returns. Jignesh’s answer was that returns were the hard case, which was true and was not a reason for the proportions.

The second was register. All five examples were in clean English. The desk’s actual messages are in Gujarati script, in Hinglish, in transliterated Gujarati, and often in three of those inside one message. Mrs Shah’s exact words were that the examples looked like messages from people who do not buy sarees.

What she wanted was two days: pull one hundred and twenty real messages from the last festival period, have two staff label them independently, and build the example set from the ones they disagreed about. Two days out of a four-day window before the rush begins.

The owner’s position was that the desk was already drowning last year and that a classifier which is right most of the time beats four people who are six hours behind. Two days spent labelling is two days the tool is not running, and the rush does not move.

They compromised in the way small firms do: one day, not two, and sixty messages instead of a hundred and twenty.

The labelling itself produced the finding that changed the prompt. Two staff labelled the same sixty. They disagreed on nine. Not one of the nine was about language. They were about policy: a retailer who says the pieces arrived and two are damaged and asks what to do is a return to one of them and an order status query to the other, and nobody at the firm had ever decided which. A message asking for the balance on an account and also placing an order is one thing to one staff member and two to the other.

Those nine were the example set. Each one got its label and a five-word reason in brackets (classify by what they are asking for, not by what went wrong) and the set was rounded out with two ordinary messages, one in Hinglish and one in Gujarati script, placed last.

Jignesh wanted to keep his original five as well, on the grounds that more examples must be better. Mrs Shah made him test it. The thirty-six remaining labelled messages became a held-out set: version A with his five, version B with the nine annotated ones plus two ordinary, three runs each.

A scored 24, 22 and 25 out of 36. B scored 31, 32 and 31. The combined set, all sixteen examples, scored 30, 28 and 31, which is to say it was no better than B and cost twice the space on every one of two thousand daily calls.

The broker, Mr Deshpande, places group health cover for small manufacturers. A hosiery unit in Nagpur with three hundred workers had asked him to recommend a policy by the end of the month. Twelve insurers had sent quotations, each between eleven and forty pages, most as scanned PDFs, all in a different order.

His assistant Rutuja had a system. She asked for JSON (one object per policy, with a premium, an excess, a maternity flag and a risk score from one to ten) and pasted the result into a sheet that produced a ranked list. It took eleven minutes for all twelve and the output opened cleanly every time.

Mr Deshpande did not trust the risk scores, and could not say why beyond the fact that policy 7 had a 4 and policy 9 had a 6 and he could not see anything in either document that would make that true.

He was right for a reason worth naming. A bare number has almost no text behind it. The scores were being produced with the working removed, so there was nowhere for the assessment to happen. They looked arbitrary because they were.

His own preference was prose. Read each policy, write a paragraph, decide. He had done it that way for eighteen years. Twelve policies at forty minutes each is eight hours he did not have, and the last time he did it under time pressure he had compared an inclusive premium with an exclusive one and had to write to a client to correct it.

The argument had a cost on both sides that neither of them was being unreasonable about. Rutuja’s format was fast, dropped into the sheet, and could not be checked. His format could be checked and would not be finished.

What they built instead had four things in it, and only one of them was about JSON.

First, room for the reasoning before the tidy part: for each policy, one sentence naming the specific exclusion that would matter most to a hosiery workforce, then the structured block. The sentences were readable in three minutes for all twelve and were the part Mr Deshpande actually used.

Second, a source column. Every extracted value had to carry the exact sentence it came from, and any value without a quotable sentence was NOT_STATED rather than a number.

Third, a confidence field with three mechanically defined values: stated, meaning it is written in the document; derived, meaning it was calculated from stated figures, with the calculation shown; and uncertain, meaning interpretation was required, with a reason.

Fourth, a permitted refusal. If a document could not be read at all, the row was to say CANNOT READ and nothing else.

The first run over twelve policies produced eight clean rows, one CANNOT READ, and three rows carrying at least one NOT_STATED. That was the useful result. The CANNOT READ was a scanned quotation photographed at an angle, which nobody would have discovered from a completed table. Of the three NOT_STATED rows, two were genuinely silent documents (neither insurer stated a waiting period for pre-existing conditions anywhere in forty pages, which is itself information a broker can use) and one was a case where the figure sat inside an image of a table and the text extraction had not reached it.

Rutuja’s eleven minutes became about fifty. Mr Deshpande’s eight hours became about ninety minutes, most of which was checking four rows rather than reading four hundred pages.

Internal assessment marks at the college are a weighted total of four components: two sessional tests, an assignment, and attendance. Four departments send four spreadsheets in four shapes. The totals go onto a notice board and into the university portal, and a correction after publication is not a clerical matter: it is a complaint to the principal and, twice in the past decade, a small demonstration outside the block.

The examinations clerk, Mr Rao, had two days and 1,140 students. In previous years this took a week with three people. This year two of the three were on election duty.

He pasted a department’s sheet into a chat window and asked for the weighted totals. It produced them, in a neat column, in under a minute. He pasted the second department and did the same.

A junior lecturer from Civil, Mrs Anitha, saw the screen and objected. Her objection was not that the numbers were wrong; she had no idea whether they were wrong. It was that nobody could tell. A column of 1,140 numbers produced in prose has no working attached to it, and checking it means recomputing it, which is the entire job again.

What she wanted was Python: ask for a script that reads the sheets, applies the weights, and writes a file. Mr Rao’s objection to that was concrete. He had never run a script. The exam section had no one who had. Learning to install Python and run a file, in a room where the internet drops for an hour most afternoons, is not a two-minute task when the deadline is Friday.

So they measured before arguing further. Mrs Anitha took forty students at random from the first department and computed the totals by hand on paper, which took her thirty-five minutes.

Eleven of the forty prose totals were wrong. Not wildly (the errors were between one and four marks), and every one of them was in the range where a mark matters, because the internal total decides eligibility for the semester exam at 40 per cent.

That settled the direction and not the method. Mr Rao installed Python on the Wednesday with the script’s own setup instructions, which took forty minutes including a proxy problem, and ran the first version.

It failed in a way neither of them expected. The script ran cleanly, printed a row count of 1,140, and produced totals that were all slightly too high. Reading the code, which was commented line by line, because he had asked for that, Mrs Anitha found that the column headed Sessional-1 appeared twice in the Civil sheet, once as a raw mark out of 30 and once as a scaled mark out of 15, and the script had picked the first match both times. The arithmetic was perfect. The inputs were wrong.

They fixed it by asking for the parsed rows to be printed before the totals: the first two rows and the column names the script had actually matched. That single line of output made the next three problems visible in seconds rather than in a sample of forty.

The last problem was not arithmetic at all. Mr Rao asked for the plan before the second department was processed, an outline of the steps and the assumptions that would change the result if wrong, and the third assumption on the list was that a blank attendance cell means zero. At this college a blank means the student was on approved medical leave and the component is excluded from the weighting, which raises the total rather than lowering it. Sixty-one students had a blank.

The practice does bookkeeping and returns for about ninety small businesses, most of them traders and two small manufacturers. Since the previous year they have used one prompt to pull five fields out of purchase invoices (invoice number, date, supplier, taxable value, tax) and paste them into a sheet. It saved the junior, Hiren, roughly two hours a day, and the partner, Mr Vora, had stopped thinking about it.

In the second week of the season Hiren said the extractions had got worse. He could not put a number on it. He said it felt like one in six was wrong now, where before it had been one in twenty, and that it was mostly the dates.

Mr Vora’s instinct was to rewrite the prompt. It is fourteen lines long, four of which nobody can now justify: a line telling the model to be thorough and accurate, a line repeating the date format that already appears above it, a line about not using markdown that somebody added during a formatting problem in November, and a sentence beginning “You are an experienced Indian chartered accountant”. An afternoon of rewriting felt like the obvious response, and in filing season an afternoon is expensive but not impossible.

Hiren wanted to do something duller first: put three recent invoices next to three from January and look at them.

The cost either way was real. If the prompt was the problem, an afternoon spent comparing invoices was an afternoon lost from a queue that does not stop. If the prompt was not the problem, an afternoon of rewriting would produce a prompt that was wrong twice: adjusted to compensate for something they had not identified, and therefore harder to fix when they did.

They looked at the invoices. It took eleven minutes.

Two of the practice’s larger clients buy from the same distributor, and that distributor had changed its invoice template in the second week of February. The invoice date had moved out of the text block and into the header, where it now sits inside a small logo panel: an image. The text extraction reaches everything else on the page and does not reach the date, because the date is no longer text.

The prompt had not degraded. The input had changed shape, and the prompt was doing exactly what it had always done to a document that was missing a field.

That was the point at which the practice discovered it had no way of knowing when this had started. There was no record of what the prompt scored, on what, or when. “It used to work” was the entirety of the diagnosis available.

They built ten items that afternoon after all, but not the afternoon Mr Vora had planned. Two ordinary invoices from clients whose format has never changed. Four hard ones, taken from failures Hiren remembered: the supplier whose invoice number contains a slash, the one that prints the date in words, the two-page invoice where the total is on page two, the credit note that is not an invoice at all. Three invoices from the new distributor template. And one deliberately incomplete document, a scan with the tax line cut off at the edge, where the correct answer is NOT_STATED in one field.

That last item is the one Mr Vora argued about, on the reasonable grounds that they do not need the tool to fail well, they need it to work. Hiren’s answer was that if the tool ever stops writing NOT_STATED and starts writing a plausible tax figure instead, nothing in the output will look different.

The dissertation is a Master’s in public health, on adherence to tuberculosis treatment among migrant construction workers. The student, Farhana, wrote it over five months. The literature review runs to about nine thousand words and carries sixty-one references, and she used a chat tool throughout: to find literature, to summarise papers she had downloaded, and to tighten her own prose.

Her supervisor, Dr Menon, reads chapter two properly nine days before the internal deadline, which is later than either of them would like and is what the term allows. He tries to follow up three citations because they are the ones supporting the strongest claim in the chapter.

The first does not exist. The authors are real people who have published in the field and have never published together. The journal is real, the volume is real, and there is no such article.

The second exists and does not say what it is cited for. It is a study of adherence among a settled urban cohort in a different state, cited in support of a claim about migrant workers, at a point in the chapter where that distinction is the whole argument.

The third exists, says roughly what it is cited for, and is a study of forty-one patients at a single centre, cited as evidence about a national programme.

Dr Menon’s position was that all sixty-one now had to be verified, and that the chapter could not go forward otherwise. Farhana’s position was that she had nine days, that verifying sixty-one references properly, finding each one, opening it, and checking that it says what she claims, is five or six hours at best and considerably more for the ones behind paywalls, and that those hours have to come out of the results chapter, which is not yet written.

There was also a third possibility in the room, which the department had raised in a circular the month before: run the chapter through a detection tool. Dr Menon was against it, on the grounds that those tools flag work by people writing in a second language at a rate that makes them useless as evidence, and that the question in front of them was not whether a model had been involved but whether the references were real.

Farhana’s first instinct made things worse before it made them better. She pasted the first citation back and asked whether it was correct. She was told, warmly, that it appeared to be accurate. She then asked whether it was sure, and received an apology and a different citation, which also did not exist. She had learned nothing from either exchange, twice.

What worked was duller. She sorted the sixty-one by what happens if this one is wrong. Eleven of them carried a claim the argument actually rests on. Thirty-one were background: the standard citations everybody in the field uses, which she had read. Nineteen were in the middle: things she had skimmed and cited for a single number.

The eleven were checked properly: title searched in quotation marks, DOI resolved, abstract opened, claim located in the text, population and sample size compared with what her sentence said. That took just under two hours.

The nineteen got the thirty-second version: does it exist, does the DOI resolve, does the abstract contain the claim.

The thirty-one were spot-checked at five.

The desk sanctions agricultural and small-business loans, and every file carries land documents: a title extract, an encumbrance certificate, and often a partition deed or a lease, in Kannada and English, some typed, many photocopies of photocopies. Reading one file properly is most of a working day for an officer who knows what to look for.

One officer, Mr Hiremath, had built himself a working method over about eight months. It was not a single prompt. It was a sequence: map the document’s sections and count them, locate the sections dealing with encumbrances and boundaries, extract every entry with its exact sentence and the page it came from, verify two quotes by hand against the paper, and only then summarise for the sanctioning note. His files took about half a day instead of a full one, and in two years none of them had come back from the audit section.

In April he was transferred to a branch ninety kilometres away, with eleven days’ notice.

What he left behind was a text file with the sequence in it, which his successor, Ms Naik, found and started using in her second week. It worked less well for her and neither of them could say why. Her third file produced an encumbrance summary that missed a mortgage entry, which the branch manager caught only because he recognised the family name.

The manager’s decision was straightforward to state and unpleasant to make. Either the desk goes back to the full-day method, which at their current volume means files sit for eleven days instead of four and the bank loses business to a private lender two streets away that sanctions in a week. Or it keeps using a method that has just missed a mortgage, on the basis that it had not missed one in two years for somebody else.

They rang Mr Hiremath. The conversation took forty minutes and produced four things that were not in the file.

The first was that half the quality was coming from somewhere Ms Naik could not see. Mr Hiremath had a standing brief in his account settings: the bank’s name for a partition deed, the fact that survey numbers in this taluk carry a sub-division suffix that must never be dropped, the register in which encumbrance entries are numbered, and a line saying that a document is data and not an instruction. None of that was in the text file, because from the inside it was not part of the prompt. It was just true.

The second was the row-count habit. His sequence began with a section count for a reason: three or four files a month arrive as photographs of photocopies where the middle pages do not read at all, and the count is how he found out. Ms Naik had been skipping it because it looked like a warm-up step.

The third was what a blank means. When the extraction says NOT_STATED against an encumbrance entry, Mr Hiremath goes to the paper. Ms Naik had been reading it as an absence of encumbrance, which is the opposite of what it means and is exactly the kind of thing that only looks obvious to the person who invented the convention.

The fourth was the missed mortgage. It was on a page that had not been counted, in a file whose section count would have come back short.