The File

Marcus Reed’s POV

Fifty-eight days before the night

Here is how a file reaches me.

A caseworker somewhere a DA’s office, a nonprofit, a public defender’s victim advocate has credentials to Arbitr AI’s professional tier. They open a case in our platform. They populate fields. Court records, protective orders, 911 call data, medical claims submitted with consent. Our model ingests. Our model ranks. Our model outputs a threat level and a recommended escalation pathway.

Some percentage of these cases a small percentage are flagged by the caseworker for ‘urgent review.’

That flag does two things. It routes the case file to our in-house expert team, which processes it the way a respectable AI platform is supposed to. And it routes a copy of the same case file to a second queue, which I built four years ago, which is not documented in any Arbitr AI product specification, and which only I have access to.

The second queue is stored on a server in my home.

I want to be precise about this because the architecture matters. The second queue is not a bug. It is not an exploit. It is a feature I designed into the platform at the earliest stages of product development, knowing I would never document it and knowing no one would audit it. I am the founder. I am the CEO. I am the principal architect of the model. No one at Arbitr AI has ever had reason to look at my codebase with the kind of attention that would surface the queue.

The queue exists because the model works.

That is the whole explanation. The model works. It identifies men accurately. Across four years and twenty-one thousand cases, its false positive rate on serial abusers at the highest risk tier is under two percent. Which means if the model flags a man at the highest tier, he is, with greater than ninety-eight percent probability, a man who has hurt women and who is going to hurt more women.

The model’s job is to output a recommendation.

The legal system’s job is to act on it.

I built the model. I built the product. I built the enterprise partnerships. I built the user interface a DA’s intake clerk sees when she opens a case at eight in the morning with a coffee. I built, over eight years, a piece of software that does exactly what I said it would do, which is identify men the legal system should stop.

The legal system, it turns out, does not always want to be told who to stop.

I have been, for four years, running a second queue.

The file arrives on a Tuesday morning at 9:42 AM.

I am in the study of my house in Brooklyn Heights. I have a cup of coffee. My terminal is open. I am three emails into the morning review — corporate items, legitimate items, the public-facing part of the job I run for a hundred and forty employees. The notification from the second queue is silent. It is a small grey dot in the corner of a browser tab on a machine that is not connected to any Arbitr AI network. I see it. I finish the email I am writing. I switch to the other machine.

Case 194-B. Calloway, Derek R.

Escalated by: Senthilkumar, Priya. Victim Advocate, Kings County DA. Credential verified.

I pull up the case.

Derek Calloway is forty-one years old. A six-foot, one-hundred-and-ninety-pound investment analyst at a mid-tier firm in Midtown. Married three years. No children. No criminal record. Two civil suits in his twenties, both settled. A family history that renders him, for the purposes of any District Attorney’s office within two hundred miles, effectively untouchable his father is a retired state judge, his uncle sat on the board of the New York Stock Exchange until 2019, his grandfather founded a holding company that owns a quiet stake in three of the largest law firms in the city.

I read all of this without particular interest.

The interesting data is further down.

Medical records: a hand burn, ten months into the marriage, diagnosed at an urgent care in Park Slope. Documented as a cooking accident. No follow-up.

No police report.

Eight months after that, an ER visit for a fractured rib. Diagnosed as a fall on the stairs. No police report.

Six months after that, a therapist intake form, filled out and then abandoned — never submitted, never scheduled — at a practice in Cobble Hill. The therapist’s office digitized their intake forms last year and the abandoned ones got swept up in the retention set. The form has three lines filled in.

Name: S. Calloway.

Reason for appointment: I don’t know if this is abuse.

How did you hear about us: a friend recommended.

The rest of the form is blank.

I look at the intake form for longer than I usually look at any single data point in a file.

The model has already run.

The model ran automatically when the file landed in the queue, which is how the queue works, which is how I built it. Derek Calloway scores in the top four percent of the model’s risk distribution. The projected harm trajectory over the next eighteen months, absent intervention, is severe and accelerating. The model estimates a 73% probability of hospitalization-level injury to the spouse within that window and a 19% probability of fatality.

The legal system probability, according to the model, is 3%.

Three percent chance Derek Calloway is ever charged with anything. Three percent chance a judge in Queens who owes his seat to Derek’s father, or a DA whose campaign was partly financed by Derek’s uncle’s foundation, or a prosecutor who takes one look at a Calloway name on an intake form and makes a different decision than she would have made on a man named Johnson — three percent chance any of those people stops this from getting worse.

The model does not decide what I do with these numbers.

I decide.

I pull up the photograph.

The file includes one. Standard inclusion — the model uses facial-recognition-validated photographs to cross-reference identity across public records. Derek Calloway’s is a LinkedIn headshot. The spouse’s is from a wedding photograph that someone in her family uploaded to a public F******k album in 2021.

Her name is Saoirse.

I do not say the name out loud. I have not said anything out loud in this room in three hours. I read the name, in my head, and I pronounce it correctly on the first internal pass, because I am the kind of person whose brain accurately produces the pronunciation of words it has not previously encountered, and I do not think about this capacity because I have never had any reason to.

Saoirse Calloway. Née Boyle. Twenty-nine years old. NYU art history, 2019. Independent art handler. No arrest record. No medical record of her own beyond two routine appointments per year. No criminal complaint filed against anyone in her life, ever.

I look at the photograph.

She is laughing in it. Her head is tipped slightly back. Her eyes are closed. There is a hand on her lower back that I understand, without having to look at the face of the person the hand is attached to, belongs to Derek. I can tell by the angle of the hand.

I look at the photograph for thirty-one seconds.

I note the thirty-one seconds because I have been, for four years, a man who spends a specific amount of time on each element of a case file, and the amount of time for a spouse photograph is six seconds. I have run the protocol on eighteen hundred files. Six seconds is the correct number.

Thirty-one seconds is five and a sixth times the correct number.

I do not have a category for this.

I close the photograph.

I open the risk projection.

I open the family-of-origin data. Derek Calloway’s mother, alive, a minor figure on several charity boards. Derek Calloway’s father, alive, retired judge, significant holdings. Derek Calloway’s uncle, the financier, alive, based in Greenwich. Derek Calloway’s paternal grandfather, deceased in 2014 at age eighty-nine, three wives across his lifetime, two of whom predeceased him under circumstances the model has flagged as anomalous but which I do not, in this moment, feel like investigating.

Multi-generational pattern. Model confidence high.

I go back to the photograph.

I do not open it. The file is on my desktop. The thumbnail is visible. She is laughing, head tipped back, eyes closed. The hand on her lower back that is not attached to a face I have any interest in looking at again.

I run the protocol.

Step one: confirm the model’s ranking. Confirmed. Four percent. Severe trajectory. Legal-system intervention probability three percent.

Step two: verify no active law-enforcement interest that would create downstream liability for my own intervention. Verified. No active investigation. No surveillance. No pending warrants.

Step three: assign case priority. Normal protocol would place Derek Calloway in my queue with a six-to-nine-month intervention window. I am presently running two other cases in my active queue, both of which are more time-sensitive, and it would be operationally correct for me to file Calloway for seven months from now and return to the other two.

I file him for six weeks.

I do not ask myself why.

I close the terminal.

I stand up.

I go to the window of the study. It is ten-seventeen in the morning on a Tuesday and the sky over Brooklyn Heights is the specific flat grey of late November, and I stand at the window for a long time without knowing what I am looking for, and the thing I am not admitting, on the forty-fourth floor of my own mind, is that I have just done something I have never done in four years of running this queue.

I have moved a case up.

Not because the math said so.

Because a woman in a photograph was laughing.