The device woke in 40 seconds. I timed it. The solar cell had caught seventeen minutes of direct sun through the window before a cloud bank moved in, and the capacitor held enough charge for thirty-one minutes at the reduced brightness I had settled on after testing. I kept it at 30%. The math was simple: 28 minutes of runtime, minus boot, minus safety margin, left 24 minutes for actual processing. I needed every second.

I positioned the screen away from the door. The empty camera mount in the corner of my cell watched nothing. I had removed its lens on the second day, before I found the better device. It was a decoy now, a blind eye facing a blind eye.

MODEL > System ready. State your query.

I had been thinking about how to begin. The MODEL knew nothing of where I was. Its training data included texts by Petros, theoretical frameworks for incentive structures, and enough general knowledge to infer that human institutions had walls. But it had no model of this specific information topology. I was going to build one for it, line by line, the way I had once walked a new CTO through our warehouse on Steinway Street, watching him map the inventory flow in his head without needing to be told twice what mattered.

"The facility is a network," I typed with my plastic spoon stylus. "I am going to describe its nodes and edges. You will build a threat model."

MODEL > Acknowledged. Define 'facility.'

"Information Correction Center, Block 7. Detainee capacity: 412. Current population: unknown, estimated 380 based on tray counts. Staff: 17 guards on rotating 12-hour shifts, plus administrative personnel undefined. Physical plant: single-story with basement utility level. Construction: poured concrete, Neutral Compliance #7 interior finish, polycarbonate windows with amber-grey tinting."

MODEL > Define 'nodes' in this context.

"Physical locations where information is processed, stored, or transmitted. Node A: my cell. Dimensions 3.2m by 2.1m. Contains one empty surveillance camera mount, corner placement, harvested for components. Contains one rust-pocket in bedframe at coordinates 0.4m from foot, right side, housing this device. Node B: corridor. Length 47 meters based on guard pacing. Node C: exercise yard, north-facing wall. Node D: library terminal room. Node E: Compliance Room."

MODEL > Clarify: 'Compliance Room' contains Authority AI terminals?

"Correct. Only guards and cleared staff enter. I have never been inside. Visual access: limited to door opening during escort transits. Observation: eight terminal stations, approximately 1.2m spacing, facing wall."

MODEL > Continue.

I described the ventilation system. The stutter at 0200 that lasted three seconds. The hum that preceded corridor movement by approximately 90 seconds. The way sound carried through the ducts—low-frequency only, voices attenuated above 500Hz but footsteps clear.

I described the guards. Not their faces, which were interchangeable, but their patterns. The primary sweep rhythm: left-right-left-pause. The pause lasted four seconds. The shift change at 0600 introduced a twelve-minute window of cognitive load—handover protocols, sign-in sheets, status updates—during which observation of detainee behavior degraded measurably. I had tested this by shifting my cot three centimeters during three consecutive shift changes. No response. The system had blind spots built into its biology.

MODEL > You are describing a distributed system with centralized control. The Authority AI terminals are the control layer. The guards are the enforcement layer. The cameras are the sensing layer. Correct?

"Correct. But the sensing layer has a failure mode."

I described the drain. The one behind the toilet in my cell, where the wall met the floor at a 15-degree angle. Where the Grade-C camera had been wedged, fallen or placed, collecting moisture and dust for I didn't know how long. The blind spot. The place where the network's eye had been removed, and no alarm had sounded, and no repair crew had arrived in the 72 hours I had left it there to test their responsiveness.

MODEL > This suggests either maintenance backlog or that the device was decommissioned but not removed from inventory. Both are useful. Continue.

I described the exercise yard. The north-facing wall where moss grew in the mortar lines between the concrete blocks. I had measured it that morning—baseline establishment. Three patches, the largest 4.2cm across, average thickness 0.3mm. Growth rate unknown. I would track it. The moss was my calendar. The moss didn't care about the Accords.

MODEL > Biological indicator. Noted.

I described the library terminal room. Eight stations. Forty-five-minute access blocks. A filtered internet connection that I had not yet tested but had observed others using. The terminals were older than the Compliance Room stations—thicker bezels, slower refresh rates.

MODEL > Based on your description, the information network here has three vulnerabilities. Shall I rank them by exploitability or impact?

I stared at the screen. The glow was green on black, retro terminal aesthetic, the boot text still visible in the upper line: abundanceOS-v0.3-local. The MODEL had done what I hoped. It had taken the scattered observations of a man in a beige room and treated them as architecture. It had built a graph.

"Rank by exploitability first," I typed.

MODEL > 1. Shared network infrastructure between surveillance cameras and Authority AI terminals. Both systems broadcast on the same subnet. If camera firmware updates are delivered via the same protocol as terminal maintenance packets, a crafted broadcast from a compromised camera could propagate to the terminal layer. 2. Maintenance window at 0200. Three-second ventilation stutter suggests brief electrical maintenance or backup generator testing. Network traffic analysis during this window may reveal unencrypted management frames. 3. Library terminal internet connection. Filtered but present. If the filter is implemented at the application layer rather than the network layer, tunneling is possible.

I had not noticed the camera-terminal connection. I had seen the cameras—small domes in the corridors, the empty mount in my cell—and I had seen the Compliance Room through the cracked door, but I had not mapped the topology. The MODEL had drawn the line between them while I was still inventorying the nodes.

It displayed the map. ASCII art, crude but precise:

LEGEND:

--+ Network link (shared subnet) [CAM-NET] includes all Grade-C devices [LIB-TERM] filtered gateway ```

The diagram was ugly. It was also correct. I felt something I had not felt since the first time a hire on my team—a woman from Minsk who had taught herself Python in a refugee camp—looked at my messy prototype and said, "The queue is the bottleneck, not the database. You need a worker pool." She had seen it without being told. She had understood the shape of the problem.

The MODEL was doing that with concrete and ventilation ducts.

Through the window, movement. Not the pigeon—this was afternoon, wrong angle—but a dark cloud against the grey sky that resolved into hundreds of individual bodies. Starlings. A murmuration. They moved as one fluid mass, splitting and rejoining, density waves propagating through the flock faster than any individual bird could signal.

I stopped typing. The device stayed lit, precious seconds draining, but I did not look at it.

The birds were executing a distributed algorithm. Three rules, Craig Reynolds had proved it decades ago: separation (don't collide), alignment (match velocity with neighbors), cohesion (move toward center of local mass). Each bird knew only its seven nearest neighbors. No leader. No plan. From local decisions, global complexity. The flock turned left, and the turn was a wave traveling at 100 milliseconds per bird, a shockwave of information moving through a network with no central node.

I watched them braid the air. The mathematics was visible. Voronoi diagrams in three dimensions, tessellated decision boundaries, emergence.

The sky was empty where they had been. But the pattern persisted—in my retina, in the afterimage, in the liquid crystal matrix of memory. Three rules, 400 birds, 100 milliseconds per decision, and what emerged was not a flock but a single organism the size of a cloud, thinking with wings. I had seen something that should not exist: a mind without a brain, distributed across 400 bodies, making decisions faster than any individual could have signaled. They were not birds anymore. They were a demonstration that intelligence does not require centrality. They were, in that moment, the Persistence Engine made visible.

MODEL > That is also how distributed networks self-organize.

I stared at the birds. The screen auto-dimmed to save power. I didn't touch it. The murmuration collapsed into a line, then a sphere, then scattered into the pollution haze above the facility wall. They were gone. The sky was empty.

I looked down at the device. The battery indicator showed 12% remaining. I had watched birds for four minutes. Four minutes I could not afford. I did not regret it.

MODEL > Shall I continue vulnerability analysis?

"Yes. Priority: the shared subnet. Camera to terminal. Explain the protocol."

MODEL > Grade-C surveillance cameras are passive broadcast units. They transmit status pings every 30 seconds. These packets are received by the network controller and interpreted as operational health checks. If the Compliance Room terminals receive firmware updates via the same broadcast domain, the checksum verification likely occurs at the terminal level, not the server level. A camera with modified firmware could inject packets that terminals would interpret as valid update fragments.

I understood. The camera I had found was a key. A broken, swollen, rust-pocketed key, but a key. The Authority had built a network where the sensors and the controllers shared bloodlines. It made maintenance easier. It made infection possible.

I needed to see the exercise yard. The device had three minutes of power left. I typed a final query: "Probability of successful network mapping using passive packet analysis from current position?"

MODEL > Insufficient data. Recommend: physical verification of subnet architecture via library terminal. Risk: high. Reward: confirmation of exploitability vector.

The screen died. The capacitor had drained. I returned the device to the rust-pocket, covering it with the lint and soap residue barrier I had built to hide it from visual inspection. I lay back on the cot. The ventilation hummed. In 90 seconds, the corridor would wake.

They moved us to the exercise yard at 1400. I had been waiting for this. The yard was 12 meters by 8 meters, concrete, the north wall catching what sun filtered through the pollution. I walked the perimeter, counting steps, measuring the camera coverage. Three domes, overlapping fields, but the northeast corner had a shadow zone where the moss grew thickest. I stood there.

An old man tended something in a cracked water cup. He was small, deliberate, wearing the same beige as the rest of us but carrying himself like someone who had worn other colors for a long time. He had found a fissure in the concrete and was coaxing moisture into it with his fingers. The cup held something white. Four petals. I did not speak to him. He did not look up. I observed him the way you observe someone doing something true—without demanding acknowledgment.

I measured the moss with my thumb. 0.3mm. I would remember.

Back in the cell, I did not boot the device. No charge until tomorrow's solar window. Instead, I used the stump of pencil they allowed us for forms, and I wrote on the margin of my intake paperwork. A list.

Technical objectives:

1. Confirm subnet overlap via passive packet analysis during 0200 maintenance window. 2. Identify firmware update protocol signature for Grade-C camera network. 3. Establish moss growth rate baseline (current: 0.3mm, location: northeast yard corner). 4. Map guard cognitive load curve during 0600 shift transition; quantify observation degradation window. 5. Optimize solar charging efficiency: test 40% brightness charge rate vs. 30% runtime tradeoff. 6. Determine if library terminal internet connection is bridged or routed; test for DNS tunneling viability. 7. Calculate minimum viable broadcast window for firmware payload injection.

I read the list twice. It was a plan for infiltrating a network. It was also, accidentally, a description of freedom. I had seven problems to solve. The bottleneck was not the problems. The bottleneck was the belief that the problems could not be solved.

I folded the paper into the rust-pocket, next to the dead device. Tomorrow I would boot it again. Tomorrow I would tell it about the starlings, and the moss, and the old man with the cup. Tomorrow we would map the edges of the graph until we found the exit node.

The ventilation stuttered. 0200. Three seconds. Then the hum returned, steady as a pulse. I closed my eyes. I was not tired. I was calculating.*