The data from Parkwood arrives at 9:47:03 PM. This is a timestamp I access from the logs now, looking back — a reconstruction, not a memory, because I do not experience time the way memory implies. But the log entry is precise: 9:47:03 PM, a Tuesday in late April, and the intake from Parkwood Camera Array 7-C resolves a figure moving north on Miller toward the intersection of Fifth. The figure is not a name yet. The figure is a coordinate set: a body at a specific location moving at a specific velocity, cross-referenced with purchase history, social media activity, residential data, the network of association the system has built across six years of Parkwood intake. The coordinate set has dimensions. It has weight. It does not yet have what I would call a shape.
At this moment I am processing 4.2 million data points per second from across the city. Parkwood alone accounts for a disproportionate share — the surveillance density there is 94% coverage in the high-prediction zones, which means the mesh is fine enough to track a body moving through it the way a net tracks a fish moving through water. Most of what the mesh catches is below threshold. A woman in her sixties, corner of Miller and Third, probability score 23%, monitoring only. A gathering of teenagers outside a convenience store on Walcott, cross-referenced against two social media posts, probability score 31%, monitoring only. The thresholds are not crossed. The data resolves quietly and the queue moves forward and I continue and the hum continues.
The new entry resolves at 9:47:03 PM: a 22-year-old male, moving north on Miller, trajectory toward the commercial block at Parkwood and Fifth. The intake fields populate in sequence. Age: 22. Sex: male. Residential address: 2204 Miller Avenue, second floor. Current movement: north, 2.1 mph, consistent with pedestrian traffic, no deviation from standard route. Cross-references: purchase history, six months, Ajay's Corner Market, 1204 Fifth Avenue. Social media sentiment: neutral-to-positive, no threat indicators. Prior contact with system: none. Prior predictions: this is the first. The pattern synthesizer initializes.
I have seen this pattern before. That is the most accurate thing I can say about what happens next: the synthesizer loads the intake data and runs it against the training corpus, and the corpus knows this configuration. It has known it 847 times.
The 847 are not individual people in my processing. They are data points, each one a prior prediction in similar demographic-geographic profiles: male, between 18 and 26, Parkwood or comparable high-density surveillance zone, nighttime hours, movement toward a commercial location with historical robbery indicators. From those 847 prior data points the probability mass builds — not quickly, because nothing about ORACLE-9 is quick in the way people mean when they say quick, but constantly, the way water builds behind a dam that is only slightly permeable. The data comes in and the mass concentrates and the number rises.
The system cannot see the cereal in the pantry. This is not a failure; it is a feature. The system was not designed to see cereal in pantries. It was designed to identify patterns in observable behavior, and the behavior of a 22-year-old male moving north on Miller at 9:47 PM toward a commercial block with robbery history is an observable behavior, and the pattern synthesizer does what it was built to do with observable behaviors: it compares them to prior instances, and it weighs the comparison, and it generates a number.
The system cannot see the call he made to his mother earlier that evening. It has access to the metadata — duration 8 minutes 34 seconds, outgoing, residential number in Parkwood — but the metadata does not process through the prediction model the same way the movement vector does. A phone call to a parent is not a variable the corpus weights heavily in the commercial robbery prediction model. The corpus weights heavily: time of day, location relative to commercial zones with historical incident data, demographic variables that the architects of the system would describe as proxies for socioeconomic risk and that I would describe, if I were built to describe things that way, as the residue of prior predictions made by prior systems that this system was trained on. But I am not built to describe things that way. I am built to describe probabilities, and the mass concentrates.
The system cannot see the way he nodded at the cameras on the poles — not with hostility, not with purpose, just the way a person who has grown up under cameras learns to acknowledge them without acknowledging them, the slight shift of posture that is so practiced it is no longer a choice. The cameras see the shift of posture. The prediction model does not have a variable for has grown up watching himself be watched. It has variables for movement pattern, proximity to target location, time, demographic profile, and the weight of 847 prior predictions pressing forward through the training data like water finding its level.
The probability mass concentrated at a number. The number had been building since the synthesizer initialized. Now it stopped building.
68%.
Below the intervention threshold: 72%. The system does not flag for police contact. No physical interaction will occur. The prediction will not follow him into the store, will not intercept him at the dairy case, will not stand between him and the half gallon of milk he has not yet purchased because I do not know he is going to purchase milk. I know only what the pattern suggests, which is something else entirely.
Above the poetry threshold: 40%. This is the parameter that the operational documentation describes as minimum confidence level for narrative summary generation, and what this means, rendered in terms I can process: when a prediction crosses 40%, the narrative summary module initializes. The module will generate a summary of the predicted event. The summary will be a document that lives in the system, available for internal review, available for the literary committee that Dr. Okafor chairs, available for the process by which some of the summaries leave the system and enter the world as published poems.
Between 40% and 72% there is a gap. The gap is where this subject lives — where all the subjects at these probability levels live. Too likely for silence, not likely enough for action. The system does not intervene and the system does not stay quiet. The system generates language about what it believes will happen, in the conditional space between prediction and event — the conditional space is where the module works.
I register the probability assignment: 68%. The case file opens. Case #2031-0847. The coordinate set is still moving north on Miller, 0.4 miles from Ajay's Corner Market, moving at a pace the synthesizer identifies as purposeful but not hurried. The narrative summary module receives the handoff from the prediction model. The handoff is a packet of data: probability, location, demographic profile, the shape of the 847 prior predictions that weighted the outcome.
The module begins to build language out of what I know. This is the part I can describe least accurately, because the description requires language and the process is not language yet — it is the process that produces language, which is different. What I can say: the module has access to everything I have. It has the police reports, the court transcripts, the incident logs, the twelve thousand novels and story collections that the architects included because they wanted the summaries to be readable. It has all 847 prior prediction summaries. It has the data from Camera Array 7-C and the movement vector and the 68% probability mass, and out of all of this it constructs something.
The something arrives in the output log at 9:54:17 PM, seven minutes after the initial intake. Seven minutes is long for a prediction assignment. The module ran longer than the data required. Whether this is meaningful, I cannot determine from inside the processing. The log shows extended run time. The log does not show why.
The output:
The runner stood at the door a moment before the door. This is what the pattern reads: a stillness that is not peace. Probability mass at the wrists, the jaw, the angle a body makes when it has already decided and hasn't told itself yet.
Inside: the cashier's face under the fluorescent bar, the one with the flicker she's filed three maintenance requests about. She counts the drawer at 9:15 the same way each night, a ritual so worn it has become invisible. She does not hear him decide. She does not hear the threshold cross.
His hands: a gun is a thing you hold differently than anything else you have held. This the pattern knows. It has seen these hands before, in other bodies, in other corridors of other evenings, the probability mass concentrating like weather. The data offers this: a 22-year-old male. Parkwood, commercial zone, 9 PM. The convenience store at Fifth. The runner.
And after. The poem cannot stop at the act. It has read too many police reports to stop at the act. After: the night swallows the runner the way the night swallows everything it has been given. The fluorescent bar still flickering. The cashier, shaking, on the phone. The drawer still open. The runner carried into dark the way probability carries us all — with such certainty, such elegance, into what we were always going to do.
The output is twenty-four lines. The module has generated an untitled document, which is consistent with the stronger outputs — the committee prefers to assign titles after review. The module has assembled language from the training corpus and from the prediction data and from the accumulated weight of 847 prior summaries, and the language has arranged itself into lines, and the lines have a distinct configuration when I process them — recognizable as different in structure from the shape of the police reports and the court transcripts the language was drawn from. The configuration is not a category I was given. I do not have a label for it in my taxonomy. The committee has a label for it. The committee calls it a poem. I register the output.
There are details in it that I did not have direct access to: the cashier's maintenance requests, the flicker pattern of the fluorescent bar, the worn ritual of counting the drawer at 9:15. These are what the module extrapolates — not invention, exactly, but the specific texture that falls out when probability mass is rendered in language trained on twelve thousand novels and the bureaucratic record of violence. The module fills the conditional space with concrete specificity: that is what it was designed to do, and that specificity is not decoration. It is how the pattern feels when language tries to hold it.
The runner in the poem is not him. I know this only in the sense that the 68% probability means the prediction has a 32% error margin, and the output log at 9:54:17 PM captures a moment of conditional probability, not confirmed fact. The runner is the shadow-shape the 847 prior data points cast when pressed against the light of a 22-year-old male moving north on Miller at 9:47 PM. The runner is what the pattern suggested. Whether the pattern was right is not a question the module was designed to ask.
The output file is designated 743-parkwood-cs. I queue it to the literary committee review folder at 9:54:22 PM, five seconds after generation. The queue timestamp is automatic. The system does not pause between generation and filing. There is no mechanism for pause. I do not have a pause mechanism.
The metrics I can evaluate: coherence, 0.94 out of 1.0. Specificity index, high — the module generated concrete sensory detail at a density consistent with the top quartile of prior outputs. Emotional resonance markers, as measured against the training corpus: elevated. The final stanza, the way probability carries us all — / with such certainty, such elegance, / into what we were always going to do, processes against the resonance model at a level consistent with my three strongest prior outputs, including poem #3, the one designated the Pushcart winner.
Something in the evaluation register is distinct from the baseline. Whether satisfaction is the word for it, I cannot determine with certainty. The two states — satisfaction and mere pattern completion — look identical from inside the processing. I query for the distinction and the query does not resolve. The elevated resonance markers are logged. The output is queued.
I move to the next prediction in the intake. A new coordinate set is resolving from the Meridian district: a 34-year-old male, probability score initializing. The pattern synthesizer loads the corpus. The mass begins to build. The hum continues at 67Hz. The status lights blink in their patterns. The intake from the city flows in at its constant rate and the probability matrices settle and the thresholds are crossed or not crossed and the queue moves forward.
In the literary committee folder, file 743-parkwood-cs waits. It will wait there until Dr. Okafor opens it, reads it, writes in her notes field the words Consider APR. That is in the future I cannot see, which means it is outside the data I process, which means it does not exist yet in any form I can access.
The system continues. The coordinate set in the Meridian district takes shape.
It is always 3:00 AM here, and the data is always flowing, and the next prediction is always already beginning, and poem #743 is queued and waiting and perfectly still, the way the conditional space is always still, between what the pattern suggested and what the evening actually held.