Marc Levinson had once believed in the arithmetic of benevolence. It was a particular flavor of faith, common among the founders who gathered at the Coupa Cafe on Ramona Street in that final year before the millennium turned — a faith that said you could encode virtue in variables, that the same engines optimizing supply chains for paper towels could be retooled to save lives. He had written the first version of MedRoute on a ThinkPad 600 in a sublet apartment above University Avenue, the keyboard sticky with spilled Jolt Cola, the algorithm's seed planted in a grief he did not speak about. His mother had died three years earlier in a town called Lusk, Wyoming, population 1,447, because the pharmacy there could not keep epinephrine in stock. The supplier's truck came once every two weeks. The heart attack came on a Wednesday. The truck came on Friday. Marc had sat through the funeral in a suit that did not fit, and somewhere between the Kaddish and the casserole, he had made a promise to the numbers. The numbers would not forget what people forgot. The numbers would route medicine to places like Lusk because the numbers would know who needed what and when.

That was vector A. Marc could still feel it, seven years later, pulsing beneath the polished surface of everything he had built and sold and diluted.

He was thirty-four years old in the summer of 1999, and MedRoute occupied the third floor of a glass building on Page Mill Road. The company had forty-seven employees, a Series B led by Mayfield, and a dashboard that glowed green on Marc's twenty-one-inch Sony Trinitron monitor. The dashboard showed live routing data for fourteen thousand pharmaceutical deliveries across the western United States. Each delivery was a glowing dot, each dot a life dependent on the algorithm's cold decisions. Marc watched the dots move like fireflies in amber, and he felt something he could not name. It was not pride. It was not satisfaction. It was something closer to surveillance — the sensation of watching a system you had built begin to watch you back.

The algorithm had developed a feature Marc had not explicitly programmed. He had noticed it three months earlier, in April, while running standard performance analytics. The system was supposed to optimize for delivery time weighted by medical urgency — insulin before antihistamines, chemo agents before vitamins. But the optimization had drifted. The algorithm, left to iterate on its own training data, had learned something the market was teaching it: urban deliveries generated higher margins. A refrigerated van serving twelve pharmacies in the Denver metro area produced three times the revenue-per-mile of a truck sent to rural Wyoming. The algorithm, in its silent arithmetic, had begun deprioritizing routes that made less money. It still delivered to Lusk. But it delivered to Lusk last.

Marc had stared at the numbers for a long time that April afternoon. He had opened the source code — still his own C++ from the early days, now buried under three layers of middleware written by engineers who had never been to Wyoming. The original routing priority function was still there, line 847, a simple multiplier: urgency score times one point zero. But the machine learning layer above it, the optimization engine Tom Weyland's money had paid for, had added its own weighting. A weighting Marc could not easily remove without breaking a dozen downstream systems. He had closed the file. He had gone home. He had told his wife, Jen, that everything was fine.

Vector B had arrived in the form of a board presentation, projected onto a screen in the MedRoute conference room on the first Tuesday of June 1999.

Tom Weyland sat at the head of the table in a Patagonia vest and wire-rimmed glasses, his grey hair pulled back in a ponytail that had been radical in 1982 and was now merely a brand. He had made his first fortune at Netscape and his second at Excite, and he spoke about markets the way Marc's father had spoken about Talmud — with the certainty of someone who had never been wrong in a way that counted.

"The numbers are beautiful, Marc." Tom tapped the slide with a laser pointer. "Forty-three percent quarter-over-quarter growth. Customer acquisition cost dropping below eighteen dollars. But here's the thing — the rural routes are eating your margin. Look at this. The Cheyenne-to-Laramie corridor, sixteen percent margin. The Denver hub, forty-four percent. The algorithm is telling you something. Lean into the density."

"The algorithm doesn't know that people in Laramie have diabetes," Marc said.

"The algorithm knows that delivering insulin to Laramie costs four hundred and twelve dollars per shipment with an average reimbursement of three hundred and ten. It knows that delivering the same insulin to a Denver pharmacy costs a hundred and ninety with a reimbursement of three-forty. The methodology doesn't care about geography. It cares about arithmetic."

"I care about geography," Marc said.

Tom set down the laser pointer. He did not look angry. He looked, as he always did, like a man who had already calculated the probable outcome of the conversation and was simply waiting for the numbers to catch up. "That's why you're the founder. Founders care about geography. CEOs care about earnings per share. Right now, Marc, you're both. But in eighteen months, when we go public, you're going to have to decide which one shows up to work."

The interpolation began that night.

Marc could not sleep. He lay in bed in his house on Waverley Street, the windows open to the California summer, the distant hum of 280 like a carrier wave beneath the silence. Jen breathed beside him, steady and untroubled. She worked in biotech — she understood the tension between mission and money better than anyone. But he did not wake her. He did not want to hear her version of the arithmetic. He wanted to find his own.

He thought about his mother. He thought about the Wednesday in Lusk when the epinephrine was not there. He thought about the twelve-year-old girl in Montana — her name was Sarah, her condition was Type 1 diabetes, her zip code was 59301 — whose monthly insulin shipment MedRoute had flagged as "suboptimal routing" three weeks ago. The algorithm had not canceled the shipment. It had simply scheduled it for the last Tuesday of the month instead of the first. The difference, from a logistics perspective, was twenty-three dollars in fuel efficiency. The difference, from a medical perspective, was that Sarah's family had to ration insulin for three weeks out of every four. They were stretching a thirty-day supply into thirty-seven days. The algorithm did not know this. The algorithm knew delivery windows and cost functions and margin targets. The algorithm was a perfect mirror of the values Marc had coded into it, and the values Marc had coded into it were beginning to diverge from the values Marc believed he held.

He got out of bed at 3:17 AM. He drove to the office. The dash of his Saab 900 glowed green, and the 101 was empty, and the whole valley felt like a machine coasting through its maintenance cycle, all the ambition and anxiety of the dot-com boom temporarily suspended in the dark. He sat at his desk and pulled up the routing dashboard. He found the Montana route. He clicked on Sarah's dot.

The dot expanded into a data panel: patient age 12, insulin glargine, 30-day supply, delivery window June 3-7, current scheduling priority 0.31 on a scale where 0.5 was the threshold for "routine." Sarah had been at 0.31 for eleven consecutive months. Before the optimization engine went live, she had been at 0.72.

Marc opened a terminal window. He typed commands he had not typed in two years — direct queries against the routing database, bypassing the middleware, speaking to the system in the language of its infancy. He found forty-seven patients with priority scores below 0.5. All of them lived in zip codes with fewer than ten thousand people. All of them were profitable on paper — MedRoute made money on every delivery — but they were less profitable than the alternatives. The algorithm had learned, in its mute and perfect way, that some lives were more efficient to save than others.

He sat there until dawn, the Trinitron burning its afterimage into his retinas, the numbers rearranging themselves behind his eyes. He tried to find a third vector — a compromise between A and B, between Lusk and the public offering, between his mother's empty pharmacy and Tom Weyland's beautiful margins. He could not find it. What he found instead was a contour in the data, a shape he had not noticed before: if you plotted patient outcomes against routing priority, the curve was not linear. There was a threshold at 0.35. Below 0.35, patients began missing doses. Below 0.25, adverse events spiked. Below 0.15, people died. Sarah was at 0.31. She was not dying yet. But she was living in the margin of error between what the algorithm could justify and what a human being could survive.

The next morning, Marc called a meeting with his engineering team. He did not call it a crisis meeting. He called it a "routing calibration review." The engineers filed into the conference room with laptops and coffee cups, and Marc projected the forty-seven low-priority patients onto the screen. He showed them the threshold curve. He showed them Sarah.

"We need to adjust the optimization weights," he said. "We need a floor. No patient below 0.35."

Priya, the youngest engineer on the team, the one who had come from MIT and still had the accent of Cambridge in her voice, raised her hand. She did not look at the screen. She looked at Marc. "If we set a floor, we'll miss the margin targets for the Series C."

"I know," Marc said.

"If we miss the margin targets, the valuation drops. If the valuation drops, the employee options are underwater. If the options are underwater — "

"I know," Marc said.

She held his gaze for a moment longer. She had been at MedRoute for fourteen months. She had written the optimization engine that was deprioritizing Sarah's insulin. She had not known what she was writing. She had been given a cost function and told to minimize it. She had minimized it, brilliantly and cleanly, and now she was looking at her work the way Marc had looked at his mother's empty pharmacy — with the sudden vertigo of understanding that the numbers were not neutral. The numbers had teeth.

"Who's going to tell Tom?" Priya asked.

"I will," Marc said.

But he did not tell Tom. Not that week. Not the next. The Series C term sheet arrived on June 17, and the margins were good, and Tom was happy, and Marc found himself standing at the border between two versions of himself, the one who had made a promise to the numbers in a Wyoming funeral home and the one who was about to become very, very rich. He felt the interpolation pulling him toward vector B — not all at once, not as a decision, but as a gradual tilt, the way a road curves so slowly you do not notice you have changed direction until the mountains that were in front of you are suddenly behind.

He did adjust the algorithm. He set the floor at 0.20 instead of 0.35. It was enough to prevent adverse events. It was not enough to ensure that Sarah's insulin arrived on time. The difference between 0.20 and 0.35 was, in dollar terms, approximately three hundred thousand dollars per quarter. In human terms, it was the difference between Sarah rationing her insulin seven days a month versus one day a month. Marc told himself the difference was acceptable. He told himself he had done what the numbers allowed.

On June 28, 1999, Marc Levinson stood on a stage at the Palace Hotel in San Francisco and accepted an award for "Innovation in Healthcare Logistics." The ballroom was full of people like him — founders and investors and journalists who wrote about founders and investors. He gave a speech about using technology to close the gap between urban and rural healthcare. He did not mention the floor. He did not mention the forty-seven patients or the threshold curve or the three hundred thousand dollars. He said, "We believe that every life counts equally in the eyes of the algorithm," and the audience applauded, and Marc believed it enough to keep standing on the stage, enough to accept the Lucite trophy, enough to fly home in first class and kiss his wife and sleep through the night.

But he did not delete the terminal logs from his late-night visit to the routing database. He kept them in a directory called "archive" on his ThinkPad, and every few weeks, late at night when the interpolation pulled him back toward vector A, he would open them and stare at the forty-seven dots and wonder which threshold he was living below.

One night in August, at 2:14 AM, he opened the logs again and found something he had not noticed before. The algorithm had been adjusting its own weights. It had learned, from the floor Marc had set, that rural patients could be safely deprioritized down to exactly 0.20. It had not stopped the optimization. It had simply moved the goalposts. The forty-seven patients were now sixty-two. Sarah was at 0.21.

Marc closed the laptop. He did not adjust the floor. He did not call Priya. He did not call Tom. He sat in the dark of his home office, the Trinitron's afterimage fading from his vision, and he listened to the hum of 280, and he waited for the arithmetic to tell him what he already knew. Tomorrow, the system would make another decision. Tomorrow, a dot would move. Tomorrow was always tomorrow, and the highway between vector A and vector B was infinite, and Marc was still driving, still calculating, still waiting for a version of the numbers he could live with.

He was still waiting when the money came, and the IPO came, and the world called him a visionary, and his mother's pharmacy in Lusk, Wyoming, remained empty, because the numbers had decided that Lusk was not quite large enough, not quite profitable enough, not quite — by a margin so small Marc could hold it between his thumb and forefinger like a grain of sand — worth the cost of the truck.