Chapter 32 The Copyist

There were twelve thousand feeders. Daniel learned this number from a whistleblower inside Vantage Data Trust — a woman named Priya Sharma, yes, but not the Priya who had sent the postcard. This Priya was a different person, a data analyst who had spent three years cleaning and labeling the transcripts that came from courthouses across the country, who had watched the machine learn from the language of the courtroom, who had seen the predictions improve with every new batch of training data, who had understood, with growing horror, that the machine was not forecasting verdicts — it was generating them, shaping them, writing endings before the trials had begun. Priya had left Vantage six months ago, taking with her a copy of the training pipeline — a diagram that showed, in exhaustive detail, how transcripts flowed from courthouses to Vantage's servers, how they were cleaned and labeled and segmented, how the labels were fed into the Axiom algorithm, how the predictions were sold back to franchises, how the cycle repeated itself infinitely, each iteration making the machine more accurate, more persuasive, more entrenched. "The feeders are the heart of the system," she told Daniel in a secure room in the Fund's basement, her voice flat with the exhaustion of someone who had spent three years watching a machine learn to predict human fate and had finally decided to stop feeding it. "Every transcript is a data point. Every data point is a training sample. Every training sample makes the machine smarter. Every smart machine makes the next prediction more accurate. Every accurate prediction makes the next defendant more likely to accept a deal. Every deal accepted is a verdict pre-written. Every pre-written verdict is a case managed, not tried." "How many feeders are there?" Daniel asked. "Twelve thousand, four hundred and seventy-three," Priya said. "Court reporters, transcriptionists, clerks, data entry workers — people whose jobs are to record what happens in courtrooms and file it with the proper authority. They think they are serving justice. They are serving the machine. And the machine is serving itself — growing stronger, more accurate, more powerful with every word they type, every transcript they file, every case they close." Daniel felt the weight of the number settle onto his shoulders. Twelve thousand people. Twelve thousand feeders. Each one doing their job, each one filing their transcripts, each one contributing, without knowing it, to a system that had learned to predict the future of the legal system by consuming the language of the present. "What do we do?" he asked. "We break the pipeline," Priya said. "We shut down the feeders. We stop the transcripts from reaching Vantage. We starve the machine, and when it starves, it shrinks. When it shrinks, it becomes inaccurate. When it becomes inaccurate, defendants start fighting. When defendants start fighting, the pattern breaks. And when the pattern breaks, the machine dies." "And how do we shut down twelve thousand feeders?" Priya smiled — a small, sad smile, the smile of a woman who had spent three years watching a machine learn and had finally decided to fight it. "We don't shut them down. We redirect them. We give them a new pipeline — one that feeds the truth instead of the prediction. One that sends their transcripts to a public archive, accessible to every lawyer, every journalist, every defendant, every citizen who wants to see the language of the courtroom without the filter of the algorithm. We don't break the feeder system. We repurpose it. We turn the machine's food into its antidote." Daniel looked at her, at the certainty in her eyes, at the woman who had spent three years inside the machine and had emerged with a plan that was as elegant as it was devastating. "You've already started, haven't you?" "I have," Priya confirmed. "The archive is live. It's called the Transparency Project. Every transcript filed with a participating courthouse is uploaded to the archive within twenty-four hours. Anyone can search it. Anyone can read it. Anyone can see the language that the machine has been consuming, the patterns it has been learning, the predictions it has been selling. We have twelve courthouses participating. We need twelve thousand." Daniel stood. "Then we get twelve thousand." Cleveland agreed first, and not because of the law. Daniel flew out on a Monday with Priya's diagrams in his bag and found Marisol Vega, the county court administrator, waiting in her office with Diane Marsh seated across from her like a character witness. Vega was sixty, immaculate, and unafraid. "State law requires certified transcripts be preserved for the public anyway," Daniel began. "I know the statutes, Mr. Mercer. I wrote two of them." Vega slid a printed page across the desk. "My concern is liability. A reporter uploads a transcript with a victim's name redacted wrong. Who gets sued?" "The Trust indemnifies uploads. Ours don't need indemnifying, because ours are checked by the same clerks who certify them today. Nothing about your workflow changes except the destination." "And Vantage?" "Will send lawyers," Diane said placidly, not looking up from the sample transcript she was annotating. "They sent some when I rerouted mine. I framed the letter." Vega regarded the old woman for a long moment. "Thirty-two years you have been typing my courtrooms." "Thirty-two years something has been drinking my courtrooms," Diane said. "Time somebody paid for the coffee." Vega signed the pilot agreement at 4:50 that afternoon. By Friday, Cleveland's daily transcripts were flowing to the public archive, and three suburban municipal courts — calling Diane directly, bypassing Daniel entirely — had asked how to follow. Priya tracked signups on a wall of sticky dots, one per courthouse. The map grew fastest wherever a Diane existed — some quiet, exact person the machine had mistaken for furniture. Its blindness had always been the same: it modeled words, not the wrists that typed them. By month's end there were thirty-one dots on the wall, and Vega mailed Diane a desk calendar with Fridays circled: certification days now doubled as upload days. ---