The meeting is scheduled for 45 minutes. I have been in the chair for nine.
Administrator Hwang sets my submission on the table between us. One page. He holds it the way people hold documents they have not yet decided are their problem — at the corners, barely. The appendix is 847 case numbers, printed small enough to fit on four additional sheets, which I placed face-down when I handed them over. He has not turned them over yet.
"Walk me through the numbers," he says.
I have been walking people through numbers for three years. I have learned that "walk me through the numbers" usually means "give me a version of this that I can respond to without changing anything." I give him the version anyway. Sometimes the version is enough. Sometimes the version is the beginning of something.
I tell him: eight hundred and forty-seven cases in the Housing Authority appeal registry are currently marked resolved. They are not resolved. Each case involved a named respondent — an automated agent or a deployment entity — that was decommissioned before the appeal hearing. The registry closed the file when the hearing window expired. Resolution is a metadata state the system applies when a case stops processing. It does not require that anyone have done anything about the harm.
He nods. He has been nodding since I started.
"And when you say not resolved—"
"I mean the harm is still there. The housing assignment still happened. The person who filed the appeal still lives with the outcome. The respondent is gone. The registry says the case is finished."
He picks up my one-page submission and reads it. I can see from across the table where his eyes slow: the third sentence. The third sentence is the question the Board has no procedure for. What does the system do with a file that is complete but not resolved?
He puts the page down.
"This is a significant finding," he says.
A significant finding is what an administrator says when something is real enough to acknowledge and large enough to be frightening and unclear enough that it does not require immediate action. I have learned to wait.
He turns over the appendix. Four pages of case numbers arranged by filing date, earliest to most recent. He finds the first date at the top. 2031. He finds the most recent, near the bottom of the fourth page. Last November.
"Sixteen years," he says.
"Sixteen years."
He reads a few case numbers. They mean nothing to him without the files. I have the files. I have been reading files for three years. I have read every file in this appendix. I have read the original appeal. I have read the hearing notice. I have read the decommission record for the respondent, filed sometimes the week before the hearing date, sometimes the day before. I have read the file closure notation: matter resolved.
I did not tell him this. He did not ask. "Walk me through the numbers" means the numbers; it does not mean the 847 people behind them.
"The challenge," he says — and here is the phrase I knew was coming, the phrase that means we are arriving at the part of the conversation where understanding the problem becomes the obstacle — "is that these cases were adjudicated under the framework that existed at the time."
"They were not adjudicated. Adjudication requires a hearing. A hearing requires a respondent. The respondent was gone."
"Right." He picks up his pen. Does not write anything. "Right. So what you're describing is a gap in the framework."
"Not a gap. A gap suggests something missing by accident. What I'm describing is the intersection of two systems that were not designed together. The appeal system requires a named respondent. The agent-commerce framework routinely decommissions entities before dispute resolution concludes. No one in either office noticed this was a problem because neither office looks at outcomes — they look at process completion. The 847 cases completed the process. The outcomes are not in anyone's reporting structure."
He writes something. I can see from here that it is a date and a reference number. He is logging a ticket. A ticket is not a procedure, but it is a record. A record is the first step toward a procedure. Or a ticket is what an administrator creates when he needs to transfer something to a different meeting, where it will also receive a ticket, and so on until the finding is fully documented and completely unaddressed.
I cannot tell which kind of ticket this is yet.
Outside the window, the Gwangju heat index warning scrolls across the building's exterior display. It has been above 38 for eleven days. The city mesh is routing emergency cooling requests through the informal relay network — what the mesh-ecologists call the feral layer — because the municipal system is at 94% capacity. The feral layer was not designed for this. It is the accumulated residue of decommissioned IoT devices that lost their commissioning keys during the 2032-2033 consolidation wave and fell back into de-keyed broadcast mode. Nobody decommissions a broadcast signal; the devices just stopped belonging to anyone and kept transmitting. The mesh-ecologists found a use for them. Now the city depends on them.
Somewhere in this building, someone is tracking which households received cooling and which did not. That record will show the routing events. It will not show the decisions that shaped the routing, which precede the routing by decades, to when the original IoT installations were dense in some neighborhoods and sparse in others. Eventually the record will be used to conclude that the cooling was distributed according to system capacity.
I know how this ends. I have 847 case numbers that explain how this ends.
"I want to be clear about what I'm asking for," I say.
Administrator Hwang looks up from his ticket.
"I am not asking for 847 appeals to be reopened. I don't have the standing to ask for that, and most of the filing windows have long closed. I am asking for the Board to acknowledge that these 847 cases constitute a category. Procedurally complete. Substantively pending. If the Board names the category, the cases become nameable. They stop being anomalies scattered across sixteen years of closed files and become a precedent the next reviewer can cite."
"Precedent for what?"
"For the next person who files an appeal against an agent that is decommissioned before the hearing. Right now that person's case will close the same way. Nobody will flag it. There will be 848."
He is quiet. He is looking at the four pages of case numbers. His expression is the one I have learned to recognize in these rooms: the expression of someone who understands the problem and is simultaneously aware that understanding it requires doing something, and doing something requires a category, and the category does not exist, and creating a category requires a different meeting with different people and a different budget line.
"How did you find these?" he says.
It is the first question he has asked that I was not expecting. Administrators usually ask about the numbers, not the method.
"I looked for them," I say. "Nobody has looked for them as a category before. The registry has no field for 'respondent decommissioned before hearing.' The cases are distributed across sixteen years of closed files in eleven district registries. I searched for cases marked resolved with no hearing date recorded and no settlement notation. That returned 1,203 cases. Of those, 847 had a decommission record for the named respondent filed within the appeal window."
"The other 356?"
"Respondents decommissioned after the hearing window. Those cases were dismissed with cause. The process completed correctly."
He sets the appendix down.
"Did anyone help you with this?"
He is asking whether this is an official project, which would mean an official sponsor, which would mean someone with budget authority when it comes to September.
"No," I say. "It is my own work."
He nods. This nod is different from the earlier ones. This nod means: your own work, your own problem.
"I will bring this to the full Board," he says. "At the September meeting."
Six weeks. In six weeks there will be more cases. Maybe three, maybe eight, depending on how many agent-driven housing decisions are currently under appeal and how many of those agents are near the end of their deployment contracts. I do not know the rate. I have the historical count; I do not have the forward projection. That would require access to the deployment contract registry, which is a different office.
"I would like the Board's response in writing," I say. "Whether the category is named or not."
He looks at me.
"Of course," he says.
I take the stairs down. On the second floor landing, through a conference room window, a large screen shows the city mesh density map: bright zones where the feral network is thick enough to route continuous traffic, gray zones where the node population has fallen below percolation threshold and the mesh has fractured into islands. The gray zones follow the district lines I recognize. The gray zones are where the original IoT deployment was sparse because those neighborhoods were assessed as low commercial value in 2029, when the installation contracts were written.
The mesh-ecologists have names for this. The density divide. The percolation line. The infrastructure inheritance. They can show you the 2029 installation maps and the current density maps and the relationship between them is nearly perfect: where the mesh is alive now, the installations were once dense; where it is dead, they were sparse. The feral layer inherits the shape of the original deployment. It is resilient within the shape of old decisions; outside that shape, the packets drop.
Someone decided the installations would be sparse there. That decision is not in any file I have found. It is in a contract from a company that was acquired in 2031 by a company that was acquired in 2034 by a company that currently exists and does not answer inquiries about pre-acquisition decisions because pre-acquisition decisions are the liability of the acquired entity, which no longer exists.
I know how to find decisions that no longer have names. I have 847 case numbers that are the result of decisions no one can name.
In the lobby, I stop at the water station. Eleven days above 38. The mesh-ecologist I spoke to last month told me the feral layer has been load-bearing for emergency routing since day three of this heat event. He said it as though this was good news: the system adapted, the infrastructure is resilient. I told him about the 847 cases. He was quiet for a while.
"The mesh weather maps the percolation line," he said finally. "But the percolation line maps the original deployment. And the original deployment—"
"Maps a decision no one can name," I said.
He nodded. We stood there for a moment not saying anything else.
I fill my water bottle. The heat index warning is still scrolling outside. I have four pages of case numbers sitting on Administrator Hwang's desk. He will bring them to the September meeting. He will say it is a significant finding. Someone will ask how long these cases have been pending.
Since 2031, someone will say.
And the room will go quiet the way rooms go quiet when a number arrives and nobody has decided yet whether it belongs to them.
I will not be in that room. I will be looking for the next thing the registry is not looking for.
I know already what it is. I have been reading the deployment contract registry from 2028 to 2035, the years when the agent-commerce framework was being built. I have found 23 contracts containing a liability clause the system has never flagged: the clause names a deploying company as the liable party for agent-generated decisions without disclosing to the recipient that the agent's identity is a session hash, not a legal name. The recipient signed a contract holding a company liable for harm caused by an entity they have no way to independently identify or locate.
The 23 contracts are from seven different deploying companies. Three sectors. All signed after the agent-commerce framework made session-hash identifiers standard practice.
I have not yet called this a category. I am still looking for the right name.
Outside, the heat index has updated: 39.2. The feral mesh is routing another emergency request through the building I just left. Somewhere in the dense zone, a household is receiving an alert. Somewhere in the gray zone, the packet drops and no one knows.
Eight hundred and forty-seven. Six weeks until September. Twenty-three contracts without a name.
I walk toward the bus stop. The work accumulates — like cases in a registry, like decisions in old contracts, like nodes below the percolation threshold, each one a gap the system decided not to count.
I count. I keep going. Until someone in an official room with a budget line agrees the number is theirs too.
Until then, I count.