The camp had a budget for 4,000 posters and one week to get them pasted. A junior officer with a laptop and a free browser editor made the first version in an afternoon, and it contained everything anyone had asked for.

Across the top: the department's logo, the state emblem, and a photograph of the district health officer. Below that, in a decorative script, "Measles and Rubella Free Nashik". Then the dates, the words "free of cost", a list of the eleven centres in two columns, a helpline number, a QR code, and a line of small text naming the scheme the funding came from. Every element was large. Several were in red, because red had been described in the meeting as urgent.

What the review meeting was about

The poster was circulated, and the comments arrived in the usual shape. The programme officer wanted the helpline bigger. The communications cell wanted the logo bigger. Somebody wanted "free of cost" in a bigger red. Nobody said the poster was hard to read, because nobody looks at a poster they have already read four times.

An intern on placement from a polytechnic did two things that cost nothing. She took a screenshot, desaturated it, and blurred it. Everything came back as one even grey field with a darker rectangle where the photograph was. Then she printed three copies, showed them to three people in the corridor for three seconds each, took them away, and asked what they had seen.

One said "a government thing". One said "something about vaccination". One named the district health officer, who she recognised. Nobody said a date. Nobody said a place.

The decision

The fix was arithmetic, and it was also political. A poster read from four metres, by a parent walking past with a child on one hip, can carry one thing first and two things second. The intern's proposal was:

• First: the dates, at roughly four times the size of everything else.
• Second: the words "free vaccination" and the name of the local centre, with a blank space left for each centre's name to be overprinted.
• Everything else at the floor: logo small in a corner, helpline small, QR code small, scheme credit small, and the photograph removed.

Removing the photograph was the part that stopped the room. It was not decoration in the ordinary sense. The photograph was there because it had been there on the last four campaign posters, and because taking it off was a statement somebody would have to make to the officer whose photograph it was.

The cost on the other side was concrete and countable. The blurred image had already shown that the photograph was the darkest, largest mass on the poster, and a face attracts fixations whatever is done to it. Leaving it in meant the strongest element on a poster about a date was a person's face. Eleven centres, two days, a target of 9,000 children, and one chance for a parent to catch the information from across a road.

The programme officer put the trade-off in a single sentence in the meeting: if the photograph stays, the date does not get read; if the date does not get read, the camp is under-attended, and the photograph is on a poster about a camp that failed.

There was a second, smaller decision hiding inside the first. The eleven centres could not all be named at readable size on one poster. Naming them all meant a list at 14 points that nobody would read at four metres. Printing eleven versions, each naming one centre, cost more in plate setup and made the print run harder to manage with a week to go.

The office had a recurring complaint that everybody treated as a fact of life: a large share of application forms arrived filled in wrongly. Phone numbers in the email row. The guardian's name in the applicant's name field. The community certificate number in the box for the transfer certificate number. Two clerks spent roughly the six weeks of the admission season telephoning applicants to correct entries, and the office had come to describe this as "the students are careless".

A lecturer who taught a first-year design elective asked for a blank form and measured it.

What the measurements said

The form had been laid out in a word processor, and every gap in it was a gap somebody had made by pressing Return. Measured, the vertical distances were 14, 16, 15, 18, 16, 22 and 14 points, in no pattern. The most important measurement was the one that repeated: a label sat about 14 points above its own box, and the next label sat about 16 points below that box.

To a reader, those two distances are the same. The label therefore belonged to both fields, or to neither, and which one a person chose depended on where their eye happened to land. The words on the form were correct. The spacing said something different, and the spacing is read first.

There were two further faults. Four groups of fields — personal, academic, community, contact — were separated by exactly the same space that separated the fields within them, so the form read as one undifferentiated list of twenty-three boxes. And the outer margin of the sheet was 8mm while the largest internal gap was 22 points, so the content looked as though it was sliding off the page.

The decision

The repair was not difficult to describe. Label 4 points above its own box, 24 points to the next field, 40 points between groups, a 16mm outer margin, and every one of those values drawn from a single scale of 4, 8, 16, 24, 40.

The problem was that the form had to remain one sheet of A4, printed both sides. It already was, at the current spacing, with the back side used for the declaration and the checklist of enclosures. Adding the space the grouping needed pushed the form to three sides, which means two sheets.

The numbers were available. The polytechnic printed about 6,000 forms a season for walk-in applicants and district outreach camps. A second sheet added roughly nine rupees per applicant in paper, printing and the heavier envelope, which came to somewhere near fifty-four thousand rupees a year, in a budget where that was a line somebody would have to defend.

Against that sat an estimate the office had never made. The two clerks reckoned they corrected about four hundred forms a season, at fifteen to twenty minutes each once the telephoning, the re-checking and the re-entry were counted. That is between one hundred and one hundred and thirty hours, twice over, in the busiest six weeks of the year, and it did not include the applicants whose forms were rejected outright and who did not call back.

The registrar's objection was not that the arithmetic was wrong. It was that the printing cost was a visible line item he would be asked about, and the clerks' hours were already paid for and therefore invisible. Two real costs, one of them countable in a budget and the other countable only if somebody chose to count it.

The trust had one designer, working on a laptop with free software, and a printer who needed the file by Thursday evening to deliver on Monday. The leaflet was A5, folded, English on one side and Hindi on the other, with the eligibility rules, the documents needed, the deadline and the trust's helpline.

The English side was set in a well-drawn geometric sans the designer had used before, at 16 points with a line height of 1.4. It looked clean. The Hindi side was set in the same typeface, at the same size, with the same line height, because that was the obvious way to make the two sides look like one leaflet.

What went wrong, and why nobody saw it

The typeface contained no Devanagari at all. When the Hindi text was pasted in, the software silently fell back to whatever the system had, which was a default Devanagari face with different proportions. Nothing warned anybody. The text rendered, the layout held, and the proof looked acceptable at the size it was viewed on screen.

A volunteer who reads Hindi looked at a printed proof and found two faults in about ninety seconds.

First, the line height. Devanagari stacks vowel signs above the headline and below the baseline, and conjunct consonants stack vertically as well, so a line occupies far more vertical space than a Latin line at the same size. At 1.4 the matras of one line were touching the headline of the line below, and in three places on the folded panel a vowel sign had been clipped by the edge of a tinted box behind the text.

Second, the letter spacing. A site-wide rule the designer used for all body text added a small positive tracking, which is correct for Latin captions and wrong here: Devanagari letters sit under a connecting headline, and adding space between them breaks the line into fragments. The volunteer's description was that the words looked as though they had been pulled apart.

Neither fault was visible to the designer, who does not read Devanagari. Both were immediately visible to somebody who does.

The decision

It was Thursday afternoon. The fixes were not large: set the Hindi in an openly licensed Devanagari family, raise the line height to about 1.7, set letter spacing explicitly to normal for that block, and re-flow the panel, which would change where the text broke and therefore the fold.

The estimate was a full working day, because re-flowing the Hindi changed the panel depth, which changed the English side's alignment, which meant re-proofing both. That pushed the file to Friday evening and delivery to Wednesday.

The cost of the delay was specific. The application window was fifteen days and had already started. Two days of the distribution plan — the district outreach van and the Saturday camp at two schools — would be lost, and the trust estimated those two days at somewhere between two and three thousand leaflets reaching families directly rather than sitting in a bundle.

The cost of shipping on Thursday was also specific, and harder to put a number on. About seventy per cent of the intended readers would read the Hindi side and not the English one. What they would receive was a leaflet with clipped vowel signs and broken words — legible with effort, and carrying the unmistakable signal that the organisation asking for their documents had not checked its own Hindi.

The trustee who had to decide put it plainly: two thousand families who do not get a leaflet, against fourteen thousand who get one that tells them the trust does not read their language.

The app was built by a four-person team in Pune for a state-funded programme. The interface followed a brand palette produced by an agency two years earlier: a strong blue for actions, near-black for primary text, and a light grey for secondary text, hints and field labels.

The grey was documented in the brand sheet as the "supporting text" colour. Checked against white, it measured 4.62 to 1. That passes the AA requirement for body text, which is 4.5 to 1, and every audit the team ran came back green.

The complaint nobody could reproduce

Field reports had a recurring line that the programme manager could not make sense of: workers said they "could not see the small writing". Nobody could reproduce it. On the team's monitors the grey was perfectly legible, and on a phone at full brightness indoors it was fine.

A developer took a phone out to a supervisor's meeting in Ahmednagar in March, at about eleven in the morning, and opened the app in the open courtyard where most of the workers were sitting. The primary text was readable. The grey was gone. Not faint — gone, indistinguishable from the white field it sat on, at the brightness the phone had settled on after a morning outside.

Three things had stacked. Phone screens in daylight lose most of the difference between a pale grey and white. The handsets in use were inexpensive Android devices with narrower gamut and poorer contrast than the team's own phones. And the audience included a substantial number of workers over fifty, whose eyes receive materially less light than a young team's.

The grey had a headroom of 0.12 above the requirement, and every one of those three conditions ate more than that.

The decision

The repair was one line in a stylesheet. The secondary text colour moved from the brand grey to a considerably darker one at about 7.4 to 1, and the field hints, which were the worst case, moved to the primary text colour with a smaller size carrying the distinction instead.

The obstacle was governance. The brand palette had been signed off by the department's communications committee, and the grey was in a document with an approval reference on it. Changing an approved colour meant a fresh note, a committee that met monthly, and — the team was told informally — a likely request to see the whole palette again, which would reopen questions about the blue that nobody wanted reopened.

The estimate was three weeks at best and a quarter at worst.

Against that sat the field cost. The hints under the fields were where the form explained what went in them: which date, whose name, what unit. A worker who cannot read the hint fills the field from memory of the training, and the commonest errors in the incoming data matched exactly the fields whose hints carried the most information. The programme was in its baseline survey, and the baseline is the number every later comparison is made against.

The team's lead put the two costs next to each other in the note she wrote: wait for the committee and collect a quarter of a baseline survey with a known data-quality fault, or change an approved colour without approval and answer for it afterwards.

The centre had grown from one room to four batches, and the two founders taught physics and chemistry between them. For the new session they paid for a photographer, a day's shoot, and a designer to produce the campaign.

The brief was to use one photograph across everything, because the budget would not stretch to three shoots and because the two founders wanted the same image on the hoarding, the feed and the printed handbill outside the railway station.

The photograph delivered was competent and had a specific problem. Both teachers stood at similar distance from the camera, at similar size, both facing the lens, one slightly left of centre and one slightly right, with the centre's board behind them. Everybody who saw it said it felt flat, and nobody could say why.

The diagnosis

The designer ran two checks. She blurred the image heavily and found two masses of almost identical weight, plus a third from a bright window behind the right-hand figure. Then she showed the composition to five people for three seconds each and asked what they had seen first.

The answers split: two said the left teacher, two said the right, one said the window. That split was the finding. Viewers disagreeing with the designer is an opinion; viewers disagreeing with each other about what they saw first is a measurement, and it says there is no primary at all.

There was a second problem underneath the first. The image had to crop to 16:9 for the hoarding, 1:1 for the feed and 9:16 for the story. With the two figures placed near the left and right thirds, the vertical crop could not contain both without cutting one at the shoulder, and the square crop put a figure hard against each edge.

The options, and what each cost

The designer put four on the table with the cost of each stated.

Crop so one teacher is cut by the frame edge. Cheapest, strongest, and available immediately from the file they already had. It makes one founder secondary and one primary, permanently, on a hoarding their students' parents drive past.

Reshoot with one teacher turned to look at the other. Produces the best image, because it converts two competitors into a relationship with a direction, and it also solves the crop, because the pair can then be composed closer together. It costs a second shoot day and pushes the campaign past the date the hoarding contract started, which was already paid for.

Blur or darken behind one figure. Subtle, survives scrutiny, available after the fact. It still leaves two figures at similar size, so the oscillation is reduced rather than removed, and it does nothing at all for the vertical crop.

Commit to an exact symmetry. Place both at identical size, identical treatment, equal distance from a central axis, looking straight out. A deliberate tie reads as a statement rather than an unresolved argument. It is the only option that treats the two founders equally, and it makes the hoarding static and the 9:16 crop nearly impossible, because an exact pair needs horizontal room the vertical frame does not have.

The decision was never really a visual one. Two people who had built the centre together had to agree which of them the campaign would be about, or agree to spend money and time so that neither had to be.

The site was well liked in the office. It opened with a full-width photograph of a new coach on a coastal road, taken by the owner's nephew, with the search form below it. On the shop's machines it loaded instantly and looked expensive.

Bookings through the site were about a fifth of what the operator had expected, and the counter staff had an explanation nobody had tested: people in Madurai prefer to book at the counter.

The morning at the stand

One of the developers spent a Tuesday morning at the bus stand with a second-hand phone he had bought for two thousand rupees, on a prepaid connection, standing in the open. He tried to book a ticket the way a passenger would.

The hero photograph was 3.4 megabytes. On the connection available at the stand, at around 1.6 megabits per second with a crowd of people on the same cell, the page showed a white rectangle for about fifteen seconds before anything appeared. Ten seconds is the limit of held attention. He tried it eleven times over the morning and it never once completed inside that.

When it did load, the photograph occupied the whole of the first screen, so the search form was below the fold and the page opened on a picture of a bus rather than on the thing the page was for.

The date field used a font size of 14 pixels, so tapping into it made the browser zoom the whole page and the layout jumped. The "Search buses" button sat at the top right of the form, which is the hardest point on a large phone to reach with one thumb — and at a bus stand, one hand is holding a bag.

The seat map was a table of six columns. At 360 pixels it scrolled sideways, which nobody discovered, so passengers thought only the first two seats were available.

Finally, the price and the departure time were set in a light grey that had looked refined in the office. In the sun, on that phone, neither was visible.

The decision

The list of repairs was short and none of them was difficult. The argument was about the photograph.

The developer proposed replacing the hero with a flat colour band carrying the operator's name and the search form, and moving the coach photograph further down the page, resized to 280 kilobytes and served at the size it was displayed. That takes the first screen from 3.4 megabytes to something under 200 kilobytes and puts the form at the top.

The owner objected, and the objection was not vanity. The competing operator's site had a large photograph. The coach was new, it was the main thing distinguishing the service from three cheaper operators on the same routes, and the owner's view was that a site that opened with a coloured band and a form would look like a government page and cost him the customers who were choosing on comfort.

Both costs were real. A page that loads in fifteen seconds loses the passengers who are standing at the stand with a bag and a departure in forty minutes. A page that shows nothing of the coach loses the passengers who are choosing between operators the night before, on a fast connection at home, where the photograph is the whole argument.

The first version of the app had four screens: the round list, a customer, a delivery entry, and a day summary. It was built in five weeks by one designer and one developer, and it worked.

The designer had read a good deal about design systems, and before the second phase began she proposed three weeks to build one: a full token set, a component library of eighteen components, a documented naming convention, and a rendered gallery. The developer was in favour. The founder asked what it would delay, and the answer was the route-planning screens, which were the reason the co-operative had commissioned the work.

The argument, stated fairly

The case for building it then was that the product was about to grow from four screens to perhaps thirty, and that every screen built without a system would have to be revisited. The designer had already seen the early signs: two card treatments that were nearly the same, three different greys for secondary text, and a button whose padding differed by two pixels between two screens because it had been drawn twice.

The case against was that nobody yet knew what the route-planning screens needed. A route screen has a map, a reorderable list, a time window per stop, and an offline state, and none of those had an equivalent in the four screens that existed. Eighteen components designed before that work would contain several that were guesses, and unpicking a guessed component after it has been used in ten places is more expensive than not having had it.

The founder's own cost was the plainest. Three weeks was a third of the remaining budget for the phase, spent on something the co-operative would never see and could not evaluate.

The compromise, and the trigger that settled it

What they agreed was narrower than either position. They spent half a day, not three weeks, on the three things that were already decided and already being got wrong: a spacing scale, a type scale and a colour ramp with a small semantic layer on top. Those three came out of decisions the design had already made, so nothing in them was a guess.

Everything else waited on a rule they wrote down: a component gets built when the same decision has been made three times, or when a second person starts making it.

A third thing was agreed at the same time and took twenty minutes: one page listing the six rules the team kept, such as one primary action per view, and space between groups at least three times the space inside them. None of those were new decisions either. They were decisions the four existing screens had already made, written down so that they did not have to be re-argued when the screens multiplied.

Both triggers fired within two months, and they fired for different things. The button reached three variants across the route screens and became a real component with a state table. The card did not — the three card-like things turned out to have different content models and stayed separate, which the designer said later she would have merged into one over-configured component if she had systematised in week one. Then a contract designer joined for the offline states, and within a week the naming convention stopped being optional, because two people were inventing names for the same roles.