On the morning the new civic map was unveiled, its streets looked complete.

Blue lines marked every road, green patches marked gardens, and small house symbols filled the neighborhoods where surveyors had walked. The map was pinned in the town hall beside a request for citizens to report missing homes. People came to point at the blank spaces. “We live here,” said Sela from the hill district. “There is no line.” “Our lane is not a road,” said an old baker. “It is a path, but people use it.”

The cartographer, Iven, unfolded the survey ledger. The map had been built from a sample: a carefully chosen group of households, streets, shops, wells, and public buildings. A sample could estimate the city’s needs. It could not remember every resident by itself. The map’s confidence was limited by the way the sample was made and by the way the answers were recorded.

Some parts were not missing because no one lived there. They were missing because the frame—the list or boundary from which the survey began—did not include them. The frame had been copied from an old tax record. The tax record had left out a workers’ camp, a new apartment block built after the last inspection, and a neighborhood whose families had repaired roofs without official permission. The map’s blank spaces were not proof of emptiness.

They were places where the survey had not looked. Iven’s assistant, Nara, put a red pin on the blank hill district. “Could we simply survey more?” she asked. “We can survey more,” Iven said, “but a bigger sample cannot correct a frame that leaves people out.”

They returned to the ledger. The first part of the survey had visited a ring of neighborhoods. A sample of that size could produce an estimate of the city’s conditions. If the households were selected randomly from a sound frame, the uncertainty from random sampling would usually shrink as more observations were added. The result would still be approximate. A confidence or margin range could describe how far an estimate might sit from the population value under the method’s assumptions.

A range was not a promise that every person lay inside it, and it did not repair a biased frame. Nara drew two circles on a blank page. The first represented the people the city wanted to know about. The second represented the people the sample could reach. A large part of the first circle lay outside the second. “A margin of error around a biased estimate can still point in the wrong direction,” she said.

The city needed a new map, but not a prettier one. It needed a better way to decide who had a chance to be counted.

They visited the camp first. A row of temporary homes stood beyond the old boundary. Families had received water and school letters there, yet no surveyor had marked the lane on the frame. Iven added the area to a new boundary. They did not simply count the homes they could see. They asked the camp council which sites existed, when people arrived, and which families were away working. Some residents had no fixed address. Others shared a shelter.

A child’s family had moved twice since the last count. The new frame still had a boundary. Every frame did. The team asked whether the boundary should follow a road, a postal route, a school district, or a map drawn from a previous survey. None was perfect. A road could pass people who worked behind warehouses. A postal route could miss a settlement without a postbox. The team chose several sources and tested whether they covered different kinds of residents.

Then came nonresponse. A hundred households were selected, but twenty-four did not answer. Some were working late. Some feared officials. Some had moved but left no forwarding address. Their answers were not a random sprinkle of noise. The households that failed to respond might share a language, a work schedule, or a distrust that shaped the result. A sample of one hundred could look tidy while missing a pattern in the twenty-four silences.

The team sent two interpreters and visited on evenings and weekends. They offered paper forms as well as door-to-door interviews. Some people answered then. The map improved, but Iven recorded every nonresponse and compared the responders with the selected households. The remaining gap was still a limit, not a blank space to fill with a guess.

Measurement entered next. A question asked whether each home had “safe water.” One surveyor wrote yes for a public pump, another wrote yes for a private well, and a third wrote no because the well was owned by a landlord. The words sounded clear until the answers were counted. The team defined the question more carefully. Safe water could mean available, nearby, affordable, and usable throughout the day. Those conditions did not always travel together.

A worn measuring cord made a street look shorter than it was. A sensor failed in one neighborhood. A clerk entered 14 where a handwritten 4 should have been. These were not all the same kind of mistake, but they could all bias the map. The team checked a sample of measurements against the original notes, trained the surveyors, and kept a correction log.

Processing came after the people had answered. A mapmaker chose which measurements counted. The group used last year’s population for a neighborhood whose population had grown, then divided a city total among streets using an old pattern. The arithmetic was flawless and the result was wrong. A program could turn uncertain observations into a clean-looking color without becoming more accurate.

Nara changed the map so viewers could see the size of the survey, the date of the records, and the places with too little information. The town hall meeting grew noisy. One speaker demanded a single exact number for every block. Iven pointed to the red pin. “A number without a sampling frame can be precise about the wrong people.”

They presented an approximate estimate for the city as a whole, with a range that showed the uncertainty from random sampling. They also listed the systematic problems: undercoverage from the old frame, nonresponse among selected households, measurement differences, and processing choices. Those problems did not vanish because the estimate had a decimal point.

A larger sample could reduce random sampling error, but it could not make a missing neighborhood appear, persuade a reluctant person to answer, repair a misread gauge, or repair a bad decision about which records to use.

Before the map was corrected, Nara climbed the hill with a notebook. She counted one lane, then found a footpath behind a row of homes. Neither appeared on the old frame. She returned with a list of names and a sketch, but the names were not counted as households until the team checked them. A note is evidence to investigate, not a number to smuggle into a ledger.

The new survey began with three lists: the old frame, a community list, and a set of places suggested by public services. The team compared them and marked disagreements. They used a random selection within the improved frame, while continuing to visit places that earlier lists had omitted. After the field work, they invited residents to mark missing symbols on a public map. The maps overlapped, but the overlap made the remaining gaps visible.

On the final day, Iven drew the hill district in green and the workers’ camp in brown. A narrow path appeared between them. He did not draw a number beneath every home. Instead, he wrote the estimate for the city, a range for the random part, and a warning beside each area where coverage was weak. The map looked less complete than the first one because it showed uncertainty instead of hiding it. Sela looked at her neighborhood for a long time. “It is here now,” she said.

Before the final map was printed, the team ran a small audit. They chose households at random from the new frame and revisited them a week later. Some answers changed: a family had received a water tank, a shop had opened, and a resident who had first refused had now become willing to speak. The estimate did not need every answer frozen forever. It needed the field process and the date beside the result. Iven added a second visit date to the ledger so a reader could see that the city was changing under the map.

Nara also tested what a larger sample could do. She drew two samples from the improved frame, one with a few homes and one with many more. The larger group usually gave a steadier estimate of the random variation, but it still missed the places outside the frame. She placed a red mark on the map where a random selection had never reached. The mark did not become less important because the total sample was larger. Systematic bias had a direction and a cause; random sampling error had a tendency to shrink with more observations from the same frame.

The town’s public meeting brought residents with lists of missing doors. Some were homes. Some were businesses. Some were people who had no fixed address but still needed water, clinics, and schools. Iven drew different symbols for each kind of service instead of reducing every need to one house icon. A map could remember more accurately when it showed the question beside each place.

The number of homes, the number of people, and the need for service were related, but they were not interchangeable.

The new map was tested against a second kind of list: water deliveries. A private cartaker had recorded the neighborhoods visited every morning, including the hill. The route showed that a hidden settlement was not hypothetical. It also revealed that the delivery company had stopped serving one court after a dispute. A sample survey could have missed both facts. The team asked the cartaker for dates and receipts, then compared the route with pump repairs and household accounts.

The sources were not identical, but they came from different activities and could reveal a coverage problem together.

Nara also revisited the households that had refused the first interview. A mother explained that the old survey had used the owner’s name, while the family used hers. The question had never been missing; the frame had been organized around the wrong person. With consent, the team recorded both names and the address everyone used for mail. That change did not prove the rest of the nonresponse was harmless.

It gave the team a concrete question to ask the other families. A confidence range could describe the survey’s estimate, while the follow-up exposed a bias the range could not represent. “And not because the map has become magically exact,” Iven replied. “It is here because people changed the frame that told the survey where to look. The next survey may find more. The map can remember that it was made by imperfect hands, and that those hands can return.”