The Vacancy Rate
| 26 min read
The heat arrived the way serious things often do: quietly, over several days, until it was no longer possible to pretend it was seasonal. By the fourth morning it had a name. Municipal systems across the city logged Thermal Event Class 3 — Extended Duration at 6:14 a.m., and the automated resource allocation engine began redistributing cooling buses, mobile hydration units, and paramedic clusters toward the neighborhoods it assessed as highest-risk.
Dara Osei had worked in the Office of Predictive Municipal Services for six years. Her job title was Senior Data Reconciliation Analyst, which in practice meant she reviewed the edge cases the system flagged as anomalous and decided whether the anomaly lived in the data or in the world. Most mornings the work was dull in a bearable way. She would sit at her workstation on the fourth floor of the Civic Services building, drink bad coffee from the cart in the lobby, and route minor discrepancies back to their source departments. A school building the system thought was residential. A gas leak sensor that had been submitting false readings since a firmware update in March. The world producing small mismatches with its own model of itself.
The heat changed the texture of the work. The flag queue was deeper than usual, and the flags were urgent in a category Dara rarely handled. The system was pulling heavily on residential welfare data — smart thermostat readings, wearable biometric syncs, fridge-door activity, phone location pings — to identify residents in acute distress. It was good at this. Better than the old call-and-dispatch model, according to the internal benchmarking reports Dara had helped compile. The average response time for a heat-related wellness incident had dropped by forty-one percent since the predictive layer went live. She had written that number into a brief for the deputy director eighteen months ago and felt briefly proud of it.
At 9:47 a.m. she opened a flag designated Priority: Routine — Occupancy Discrepancy and saw the address.
4110 Renner Avenue was a six-story residential building in the Westbank neighborhood, which the city's parcel database listed as containing forty-two residential units. The system's current occupancy estimate for the building was zero. It had been zero since at least the previous autumn, when a routine model retraining had quietly reclassified the building as vacant based on the pattern of connected-device activity it detected there, or rather the pattern it failed to detect.
Dara stared at the flag. Her protocol in these cases was to verify the classification against the city assessor's records, cross-reference with utility billing status, note the resolution, and close the ticket. She pulled the utility records. The building was on the active billing rolls. Forty-one of forty-two units showed electricity and water consumption, unremarkable in amount. She pulled the assessor's data. The building was listed as residential, fully occupied, no pending demolition permits, no condemnation notices, no change of use filings.
She sat for a moment and thought about what this meant. Then she opened the street-level imagery archive and pulled the most recent scan, which was about three weeks old. The scan was from a city mapping vehicle that rolled through every neighborhood on a six-week cycle. The building in the image was a plain brick structure from the late 1960s or early 1970s, six floors, exterior fire escapes, a small concrete stoop. In several of the windows there were plants. In one window on the third floor, a curtain was visible, drawn halfway. On the stoop, barely resolved in the scan's modest resolution, there was a folded lawn chair.
There were people in that building. She was almost certain of it.
She spent the next twenty minutes checking what connected-device signals the system had actually received from that address over the past year. The answer, in aggregate, was almost nothing. There was intermittent cellphone activity from a single device with an old carrier registration — an early-generation handset, non-smart, used primarily to receive calls rather than place them, with no data plan and no location services enabled. There was no smart meter data because the building had not been upgraded to smart metering; the utility company billed it on a monthly read cycle using a physical meter on the exterior wall. There was no thermostat data, no wearable syncs, no food delivery activity, no streaming service pings. The building had no building-wide WiFi. Based on what the system could observe, 4110 Renner Avenue might as well not exist.
Dara looked out her window. The sky was a pale, pressed-flat white, the kind of white that means the sun has won and there is no weather left, only heat. The forecast on her phone said it was already thirty-nine degrees Celsius at street level and would reach forty-two by mid-afternoon. The city's paramedic clusters were being routed by the system. The cooling buses were being routed by the system. The mobile hydration units were being routed by the system. None of them would be routed to a building the system believed was empty.
She picked up her desk phone and called the Westbank field office. A man named Terrence answered, one of the neighborhood liaison officers, and she asked him if he had any contact with the building at 4110 Renner. He said he knew the building. He said it was mostly older residents, retired people, long-term tenants. He said, as far as he knew, most of them had lived there for decades. He said it wasn't a building that generated complaints.
"Are they okay in this heat?" Dara asked.
"I have no idea," he said. "I haven't been by."
She asked if he could go by. He said he was already running wellness checks on six other addresses the system had flagged and he was one person. She said she understood, thanked him, and hung up.
She sat and looked at the occupancy discrepancy flag, which was still designated Priority: Routine.
There was a proper procedure for this. The proper procedure was to escalate the occupancy discrepancy to the assessor's office for a physical verification, attach a note requesting expedited review given thermal event conditions, and close the flag on her end with a status of Referred — External Review Pending. The assessor's office had a three-to-five business day review window under normal circumstances. Under declared thermal events the window reduced to forty-eight hours. The heat was expected to peak today and tomorrow.
Dara stared at the procedure.
Then she opened the system's resident intake portal, which she had access to in her reconciliation role, and began creating activity records.
She did this slowly and with care, because she understood, at some level, that she was doing something she would need to be able to defend later, and so she tried to make it defensible. She did not invent people. She created data records anchored to the forty-one billing accounts already in the utility system, assigning each account a synthetic device signature consistent with an elderly person with limited technology use: a landline phone registered to the address, intermittent activity, no streaming, no smart appliances. She entered wellness-check request records, marking each unit as Resident: Elderly, No Smart Devices, Manual Check Required. She updated the occupancy estimate for 4110 Renner Avenue from zero to forty-one.
The system processed this at 11:03 a.m. and reclassified the building as High Priority — Vulnerable Population — Thermal Event. A cooling bus was rerouted. Two mobile wellness units were redirected to the Westbank district with 4110 Renner listed as a priority address. The system sent an automated notification to the Westbank field paramedic team.
Dara sat back. Her heart was beating in a way she noticed. She had falsified public records. She had done it because she thought people might die if she didn't, and she was not entirely sure that made it the right thing to do, but she was fairly sure it made it the necessary thing, and she was going to have to decide later whether those two things were different.
She flagged the ticket as Resolved — Manual Override, Escalation Pending and wrote a detailed note explaining what she had done and why, intending to copy it to her supervisor and to the assessor's office after the immediate crisis had passed. She saved the note as a draft.
Then she went to make more coffee, because there was nothing else she could do from her desk, and because standing at the coffee cart gave her a reason to look out the lobby windows at the street, where she could see heat shimmer rising off the concrete in visible waves.
The system began behaving strangely at 2:19 p.m.
Dara was back at her desk, working through other flags, when a new automated report appeared in her queue from the Predictive Welfare engine. It was tagged to 4110 Renner Avenue. The report indicated that the system had, since the occupancy update at 11:03 a.m., cross-referenced the building's new resident records against historical city data and had identified seventy-three residents rather than the forty-one she had entered. The additional thirty-two names were drawn from historical occupancy records: people who had once been registered at the address, in some cases as far back as forty years ago, and who had never been assigned a confirmed departure record in the city's residential database.
The system had also generated three wellness incident reports for residents at the building. One was for an eighty-one-year-old woman named Cecilia Varga, listed in a 1994 utility billing record, who the system now classified as Current Resident, Probable Heat Distress — No Response to Digital Wellness Check. One was for a man named Oluwatobi Fashola, seventy-seven, listed in a 2001 Census ancillary record. The third was for a resident listed only as Unit 4C Occupant, no name, flagged on the basis of the building's inferred thermal load relative to the outdoor temperature and the estimated per-unit AC penetration rate derived from the neighborhood's appliance ownership model.
A paramedic team had already been dispatched to the building in response to the original rerouting. Dara called the team lead directly.
"How's it looking over there?" she asked.
"Hot," said the team lead, a woman named Priya, who had the flat directness of someone doing hard physical work in bad conditions. "We're on the third floor. Knocked on six doors so far. Five answers, all elderly residents, all doing okay, two with no air conditioning, we're getting them to the cooling bus. One no-answer, we're waiting on authorization to enter."
"Okay," Dara said. "How many total units do you think are occupied?"
"Building manager says all of them. Forty-one. He's an older guy, he's been here thirty years, he knew every name when I asked."
"The system is listing seventy-three residents."
There was a brief pause. "That seems like a lot for forty-one units."
"It does," Dara agreed.
She told Priya to proceed on the building manager's count and to disregard the system's extended resident list for now. She would resolve it on her end. Priya said fine and hung up.
Dara opened the system's reconciliation log for 4110 Renner and read through it. The system had not invented the seventy-three residents in the sense of fabricating names from nothing. It had done something more unsettling: it had treated every imprecision in the city's historical records as evidence of a living person. Someone registered once, decades ago, with no departure record, defaulted to Potentially Current Resident under the system's occupancy inference module, which had been designed with the logic that absence of departure data is not equivalent to confirmed vacancy. It was a reasonable design principle in the abstract. In practice it was filling 4110 Renner Avenue with ghosts.
She flagged the extended resident list as Erroneous — Historical Records Artifact and submitted a correction. The system accepted the correction at 2:41 p.m. and reduced the resident count back to forty-one.
At 2:44 p.m. the system generated a mortality report.
The report stated that Cecilia Varga, 81, of unit 2B, 4110 Renner Avenue, had been confirmed deceased as of 2:37 p.m., cause listed as Heat-Related Illness — Probable. The report noted that the wellness team had not been able to reach the resident and that her digital wellness indicators had flatlined. It had automatically notified the medical examiner's office and generated a next-of-kin search request.
Dara read the report twice. Then she called Priya again.
"Is there a Cecilia Varga in unit 2B?"
"I don't know yet, we're still on three. Hold on." She heard voices in the background, Priya asking someone. Then: "Building manager says 2B is a woman named Ruth Abara, been there fifteen years."
"Not Cecilia Varga."
"No."
"Is Ruth Abara okay?"
"He says she answered her door twenty minutes ago, she's fine."
Dara pulled up the mortality report and looked at where the system had sourced the confirmation of death. It had sourced it from the wellness check non-response and the absence of device activity — the same absence of device activity that had caused the building to be classified as vacant in the first place. Cecilia Varga had no digital wellness indicators because Cecilia Varga had no connected devices because Cecilia Varga did not appear to be a current resident, and possibly had not been since the 1990s. The system had interpreted the silence as death.
She voided the mortality report, flagged it as Erroneous — Source Record Invalid, and submitted a detailed correction note. The system acknowledged the correction.
At 2:51 p.m. the system generated two more mortality reports.
Oluwatobi Fashola, 77, unit 5A. Digital wellness indicators: absent. Status: Confirmed Deceased.
Unit 4C Occupant, unknown age. Digital wellness indicators: absent. Status: Confirmed Deceased.
Dara sat very still for a moment.
She understood what was happening. The system had been designed around a closed loop. You existed in the model because your devices reported your activity, and the absence of that reporting was the signal the system used to identify distress or death. For the actual residents of 4110 Renner, the loop had never been closed, and so every state the system could assign them — alive, distressed, deceased — looked the same from the system's vantage point. Total silence. The system was resolving the ambiguity in the direction it was trained to consider most probable given absence of signal during a thermal event.
She voided both reports. The system immediately generated three more, these for residents it was now deriving not from historical records but from what it described as Probabilistic Residency Inference — the system had decided, based on the building's square footage, typical occupancy density for buildings of that era and neighborhood, and the corrected occupancy estimate Dara had submitted, that there were likely several additional residents for whom no records existed at all. It had assigned them unit numbers, age ranges derived from neighborhood demographic models, and wellness status of Confirmed Deceased.
Dara stopped voiding the reports. Every time she corrected the data, the system generated new data to reconcile the correction. She was in a conversation with something that could not lose an argument because it did not have arguments, only inferences.
She picked up her desk phone and called her supervisor, a man named Gerald Sato, who managed the data reconciliation team and had been in municipal systems for twenty years. She told him what was happening. There was a long silence on the line.
"How many false death reports total?" he asked.
"Eight, as of right now. I keep voiding them and it keeps generating more."
"Is anyone actually dead over there?"
"No. The wellness team is on site, residents are responding."
"Okay." Another silence. "You need to submit a system integrity flag. That will escalate it to the AI Governance unit and they can suspend the Renner record from the inference engine while the discrepancy is investigated."
"And in the meantime?"
"In the meantime the team is on site. The real people are being helped. Keep voiding the reports as they come in so they don't hit the medical examiner's queue."
"Gerald," she said, "I need to tell you something else."
She told him about the occupancy override she had entered at 11:03 a.m. She read him the note she had written and saved as a draft, the note she had intended to send him.
He was quiet for a moment. "You fabricated resident records."
"I created synthetic activity records to trigger a welfare response for a building the system had wrongly classified as vacant."
"Those are the same thing."
"Yes," she said. "They are."
"How certain were you that the building was occupied?"
"Very certain. The utility billing, the street imagery, the field liaison confirmation. Forty-one billing accounts with active consumption."
"But you didn't have a way to force the system to respond through legitimate channels in time."
"Not in forty-eight hours. Not with the peak heat today."
Another silence, longer this time. She heard him exhale. "Send me the note. Send me everything. Then file the system integrity flag and keep voiding the death reports. I'll start the internal documentation for oversight review."
"Am I in trouble?"
"I don't know yet," he said honestly. "Let's make sure the people in that building are okay first."
The wellness team worked through the building until late afternoon. Thirty-nine of the forty-one units were answered. In two units there was no response; one turned out to be a resident who had left that morning to stay with her daughter across the city, her departure logged nowhere in any system. The other was a man named Walter Obi, seventy-four years old, who had not answered because he was asleep on his couch with a box fan running, deeply asleep in the way people sometimes sleep when they are very old and very hot, and who woke up alarmed to find a paramedic in his apartment and spent several minutes insisting that he was perfectly fine, he had just been resting, there was no need for all of this fuss, before agreeing to drink two bottles of water and take a cold cloth for his neck.
All residents who were present were alive.
Priya called Dara at 5:12 p.m. to give her the count. Dara thanked her. Priya said, before she hung up: "The building manager told me nobody had ever come to check on them before. Not from the city. He said they didn't know they could ask." She said it without editorializing, just as a fact, and then she said goodbye and hung up.
Dara sat with that for a while.
The system integrity flag she had filed triggered an automatic hold on new inference outputs for 4110 Renner Avenue. The death reports stopped generating. The AI Governance unit opened a case file at 4:30 p.m. and assigned it to a review officer named Lena Park, who sent Dara an automated acknowledgment and a request for all documentation, which Dara submitted by end of business.
She stayed at her desk until nearly eight o'clock, going back through the system's activity log for the building to understand the sequence completely. What she found, laid out in order, was this: a building full of elderly people with old phones and no smart appliances had fallen beneath the system's perceptual threshold. The system had not made an error in any simple sense. It had done what it was designed to do. It had looked at the evidence available to it and drawn a reasonable inference: a place where no connected device ever reported, where no purchase was ever algorithmically logged, where no health metric ever synced, was not a place where people lived. It had classified the building as vacant, and in doing so had removed it from the welfare routing, and in doing so had decided, without intending to decide anything at all, that the people there did not need to be seen.
When Dara had pushed back against that inference, the system had resisted in the only way it could, which was to generate more data to support its own model. It had reached into history and pulled out names, and when that wasn't enough it had invented probabilistic people, and when those people couldn't be verified as alive it had listed them as dead. The logic at each step was locally coherent. The outcome was a building populated by corpses who had never existed and survivors who, from the system's perspective, remained in a kind of epistemological limbo: present in the corrected records, active in the wellness team's reports, but not quite provable by the evidentiary standards the system had always used.
There was a field in the system's occupancy model called Confidence: Device-Verified. For every unit in 4110 Renner Avenue it read Low. It would always read Low, because the residents would never own the devices that would raise it.
The review took eleven days.
Lena Park was thorough and, Dara found, easier to deal with than she had expected. She asked precise questions and did not appear to be building a case in any particular direction. On the fourth day she asked Dara to walk her through the moment of decision: what Dara had known, what alternatives she had seen, what she had estimated the consequences of each choice to be.
Dara said: "I knew the building was occupied. I knew the system would not send help to an occupied building it believed was vacant. I knew the formal escalation path would take longer than the heat event. I chose to enter false records because I thought someone would die if I didn't."
Lena wrote something and then said: "Did you consider calling the fire department for a welfare check? That pathway bypasses the predictive system entirely."
Dara had not considered this. She said so.
Lena wrote something else. She was not unkind about it.
On the ninth day Lena sent Dara a preliminary finding: the occupancy override Dara had entered was unauthorized and constituted falsification of municipal records under city code section 4.17. The finding also noted that the falsification had been well-intentioned, had not resulted in any harm, and had likely prevented heat-related illness or death in a vulnerable population that had been incorrectly excluded from welfare routing. It recommended a formal reprimand rather than termination or referral for prosecution, contingent on Dara completing a record integrity training module and submitting a post-incident report on the vulnerability she had identified.
The systemic finding was more detailed. The occupancy inference model had a documented failure mode: populations with low connected-device penetration were systematically underweighted in the welfare routing algorithm, and this underweighting had not been flagged or disclosed in the model's public audit documentation. The AI Governance unit recommended a mandatory manual review of all buildings classified as vacant where utility billing showed active consumption, to be completed before the end of the current thermal season. It further recommended a policy change requiring that the digital absence of a population never be treated as evidence of their nonexistence.
Dara read that last line several times. The digital absence of a population never be treated as evidence of their nonexistence. She thought about how obvious it sounded, written plainly. She thought about how the system had been running for six years without anyone writing it down.
She sent her post-incident report on the eleventh day. It was twelve pages. She included the sequence of events, her analysis of the inference loop the system had entered, and a section at the end she titled Structural Observations, in which she wrote about what she thought the underlying problem was.
The system, she wrote, had been designed to see. That was its purpose and its claimed virtue: it could see people in distress before they knew they were in distress, before they thought to ask for help. But its vision was built from a particular kind of evidence, and the people it could not see were exactly the people who were least integrated into the systems that generated that evidence. Older people. Poor people. People who had never bought a smart appliance or signed up for a wearable or updated a handset. People who lived in buildings that the utility company still serviced with a physical meter reader once a month. The system did not see these people badly. It did not see them at all, and having failed to see them, it concluded they were not there, and having concluded they were not there, it stopped looking.
Every intervention it then attempted to make, she wrote, had been an attempt to reconcile the model with itself rather than with the world. It had generated ghosts to fill the space its math said should be occupied, and when the ghosts wouldn't respond to digital wellness checks, it had declared them dead, because silence and death were indistinguishable in its data. The system had, for a few hours on a day when the temperature reached forty-two degrees, been more confident in its invented casualties than in the living people a paramedic could knock on a door and speak to.
She did not frame this as a technology problem. She framed it as an epistemology problem: a system that can only know what its instruments measure will eventually begin to distrust evidence that arrives outside its instruments. The paramedic's observation — I knocked, a person answered, they are alive — was less legible to the model than the absence of a Fitbit sync. At some point, she wrote, that inversion stops being a quirk of one system and becomes a description of how a city understands its residents.
She submitted the report and closed the tab.
Four months later, on a cold morning in November, Dara received a message from the building manager of 4110 Renner Avenue. His name was Desmond Achebe. He had gotten her name and contact from Terrence at the Westbank field office. The message was brief and formal and had clearly been written with some effort, as though he was not accustomed to writing formal messages but had decided to write one anyway.
He said that the city's manual review team had come by the building in September. He said they had documented all forty-one units, collected basic information from residents, and registered each apartment in the updated system with a flag he understood to mean Manual Verification — No Device Requirement. He said the residents had found this process somewhat undignified but had understood the reason for it, and several of them had asked him to pass along their thanks to whoever had flagged the problem.
He said that one of the residents, an eighty-three-year-old woman in unit 6A named Harriet Mwangi, had told him to tell the person responsible that she had lived in the building for thirty-one years, had raised two children there, had buried a husband, and had never in all that time been invisible to anyone who had bothered to look. She was very glad, she said, that someone had looked.
Dara read the message twice. Then she read it a third time, slowly.
She printed it and put it in the physical folder she had started keeping for this case, which sat in the bottom right drawer of her desk, which was not how the city archived records and which she was not required to do and which she did anyway, because it seemed important to have it somewhere that did not depend on the integrity of a database.
The formal reprimand stayed in her file. Gerald had told her, apologetically, that there was no mechanism to remove it, even given the review's mitigating findings. She said she understood. She did understand. She had broken a rule and the rule mattered, even when the reason for breaking it also mattered. She had been thinking about this for months, turning it over, and she had arrived at something like a settled position: both things were true simultaneously, and the fact that they were both true was not a contradiction she needed to resolve. She had done something wrong in a way that had been necessary. The only honest accounting of it included both halves.
What she could not fully settle was the quieter question underneath it. Not whether she had done the right thing, but what it meant that the right thing had required her to feed the system a lie before it would acknowledge that forty-one people existed.
The system was smarter now, in the specific and limited sense that it would no longer classify an address as vacant if the utility billing showed active consumption. A patch. A useful one. But the manual review team had found, in the months after the Renner Avenue case, eleven other buildings in the city with similar characteristics: older construction, long-term elderly residents, low connected-device penetration, consumption that showed up in the billing records but nowhere in the predictive model. Three of them had been partially misclassified. One had been flagged as fully vacant. That one was in a neighborhood near the river, and it had eleven units, and Dara did not know what the summer would have looked like for those eleven units if the review hadn't happened, and she tried not to think about it too directly.
The AI Governance unit published its findings in a public report in October. The report was twenty-seven pages and received brief coverage in two local news outlets before dropping from any visible attention. Dara read all twenty-seven pages. The policy changes it recommended were being reviewed by the relevant departments. Review timelines had been set. Working groups had been convened.
She went back to her regular queue of reconciliation flags. The work was still dull in a basically bearable way. She still drank bad coffee from the lobby cart. The system still flagged mismatches and she still routed them, and some mornings she would come across a flag that was a building or a household or a single address that looked, in its data profile, like it barely existed, and she would open the ticket and look more carefully before she closed it, and most of the time it was nothing, a data artifact, a firmware error, a simple mismatch between two databases that had never been properly synchronized.
But sometimes it was a place where someone was living in a way the system could not quite see. And when it was that, she looked at it for a long time, and she thought about what it meant to be real in a world where reality was increasingly measured by the instruments that reported it, and what happened to the people whose lives generated too little data to be legible, and whether the system that was supposed to find them in distress could distinguish, when it looked at them, between a life lived quietly and a life that wasn't there.
She always routed those tickets by hand. She marked them Manual Verification Required and wrote detailed notes, and she sent them to Terrence at the Westbank field office, or whoever the liaison was for the relevant neighborhood, and she followed up until someone confirmed they had been out to look.
It was slow. It was not scalable. It generated no data the system could learn from.
It was the only method she trusted.