1. The Welfare Ghost

Natsuki Mori had stopped believing in numbers the morning they told her mother was no longer eligible.

That was 2023, six months before the first cut. The notice came on pale green paper, the kind that smelled of cheap ink and official indifference. Her mother had been on the Livelihood Protection Program for eleven years, ever since the factory closed and the doctors found the tumor in her left lung. The algorithm—new that spring, piloted in Fukuoka Prefecture—had recalculated her "minimum subsistence score" based on electricity usage, grocery store loyalty points, and the frequency of her visits to the public clinic. It concluded that she was spending 4.2% more on non-essential goods than her peer group, and therefore her standard living allowance should be reduced by 7.8%.

Natsuki had been a caseworker then, freshly graduated from Kyushu University with a degree in Social Policy. She had sat across from her mother's social worker, a tired woman named Shimizu who smelled of cigarettes and regret, and watched her circle numbers on a printout. "It's not personal," Shimizu had said. "The system just sees patterns."

Her mother died fourteen months later. Not from the tumor—that had been stable—but from the pneumonia she caught because she turned off the heater to save money. Natsuki quit her job the same week and took a position at a private data firm called Kairo Analytics. Kairo had won the prefectural contract to "optimize" the welfare algorithm. She told herself she would learn the machine from the inside, find its flaws, expose them. She told herself many things.

Now it was August 2027, and she was sitting in a windowless room on the seventh floor of Kairo's Fukuoka tower, watching a different algorithm eat a different life.

The screen glowed with a cascade of hexadecimal strings, each one tagged with a pseudo-anonymized identifier. Kairo called them "care units"—a euphemism that made the welfare recipients sound like parcels in a sorting facility. Natsuki's job was to validate the output: compare the algorithm's recommended benefit adjustments against actual caseworker reports, flag anomalies, and suggest corrections. It was tedious, meticulous work, and she had been doing it for twenty-seven months.

But three weeks ago, she had noticed something strange.

Care unit #2027-0419—a 74-year-old widower in Hakata Ward—had been flagged for a 5.2% increase due to rising rent. That was routine. What was not routine was that his file had also been cross-referenced with a separate, encrypted data stream labeled "Project Chrysanthemum." The label was not part of the standard dashboard. Natsuki had only seen it because of a rendering glitch that briefly exposed the underlying JSON structure. She had copied the tag before the system reset.

She had searched the internal database for "Chrysanthemum" and found nothing. She had asked her supervisor, a gaunt man named Kuroda with a nervous twitch in his left eyelid, and he had told her it was a deprecated testing module. "Forget it," he had said. "Focus on your validation queue."

She did not forget it.

Over the following nights, she wrote a silent script that crawled the Kairo server logs for any mention of the Chrysanthemum tag. The script returned 412 instances over the past eighteen months. Each instance involved a care unit that had received a benefit adjustment, but the adjustment was always accompanied by a second, parallel output—a set of coordinates, a time stamp, and a three-digit "affinity score." The affinity scores ranged from 0.78 to 0.94. They did not correlate with income, age, health status, or any other welfare variable Natsuki could identify.

They correlated, she eventually realized, with surnames.

Care units whose surnames matched those of prominent regional families—Otomo, Kuroda, Kim—had higher affinity scores. Not all of them, but a specific subset: the elderly, the infirm, the ones who had been on welfare for more than a decade. The ones who had once been wealthy, before the bubble burst, before the shipping conglomerates collapsed, before the textile mills moved to Vietnam.

Natsuki had grown up in a small apartment in Yoshizuka, a district that smelled of grilled mackerel and laundered uniforms. She knew the old families only from newspaper archives and the whispered stories her grandmother told about the "good years"—the 1980s, when Fukuoka had been a gateway to Asia, when Korean and Japanese merchants had traded silks and ceramics and the city had glittered with neon and ambition. Those families had lost everything in the 1997 financial crisis, and most of their descendants had trickled into the welfare system like water finding its level.

She had never imagined they would become data points in a private algorithm.

On the evening of August 17, Natsuki finally traced the Chrysanthemum outputs to an external server in Busan, South Korea. The server was registered to a shell company called Hwang Trading, which, according to a quick registry search, was owned by a holding firm in Panama. The data flow was encrypted, but the metadata revealed one consistent pattern: every time a Chrysanthemum-tagged care unit received a benefit increase, the external server pinged back a confirmation within 48 hours. And within 72 hours of that confirmation, an auction house in Osaka or Seoul or Tokyo would list a new antique—a Joseon celadon vase, an Edo-period lacquer box, a Goryeo bronze mirror—with a provenance that traced back to that same family surname.

Natsuki had printed out the correlations on a sheet of thermal paper, the kind that fades if left in sunlight. She had folded it into her jacket pocket and walked to the 24-hour convenience store on the corner, bought a can of cold coffee, and sat on the curb watching the traffic lights cycle red to green to red.

She was not a detective. She was not a journalist. She was a data validator with a dead mother and a burning suspicion that the welfare algorithm was not a tool for social justice but a stethoscope pressed against the chest of the city's old money, listening for the heartbeat of inherited guilt.

She did not sleep that night.

The next morning, she found Supervisor Kuroda dead in his office.

The official cause was myocardial infarction. The police report cited a history of hypertension and stress. But Natsuki noticed, when she was led into the room to identify his personal effects, that his right hand was clenched around a small ceramic fragment—pale green, with a subtle crackle glaze, the kind of shard that might have broken off a Goryeo celadon bowl. The officers had already bagged it as evidence, but Natsuki recognized the color. She had seen it in the auction listings that followed the Chrysanthemum pings.

She said nothing.

She collected Kuroda's belongings: his reading glasses, his wristwatch, a half-empty bottle of antacids, and a notebook. The notebook was plain black, the kind sold in bulk at office supply stores. She did not open it until she was back in her cubicle, the blinds drawn, the air conditioner humming its indifferent tune.

The notebook contained 47 pages of handwritten notes, all in Kuroda's cramped, sloping script. The first 30 pages were mundane—meeting reminders, grocery lists, a draft of a performance review for a junior analyst. But the last 17 pages were different. They were filled with dates, names, and three-letter codes that matched the Chrysanthemum tags. And on the final page, Kuroda had written a single sentence, underlined three times:

"They are not predicting poverty. They are predicting greed."

Natsuki closed the notebook and placed it in her desk drawer. She locked the drawer. Then she opened her terminal and navigated to the Kairo server's root directory, bypassing her usual validation interface using a backdoor she had discovered in her second month—a privilege escalation flaw that the IT team had never patched. She knew it was a violation of her employment contract. She knew it could get her fired, sued, or worse.

She did it anyway.

The Chrysanthemum module was not deprecated. It was hidden under a subdirectory named "archived_legacy," which was a lie. Inside, she found a real-time data pipeline that ingested three streams: (1) welfare benefit adjustment records from all nine prefectures that used Kairo's software, (2) online auction transaction histories from six major Asian platforms, and (3) a proprietary database called "Keiretsu_Mapping"—a genealogical tree of 212 families, cross-referenced with bankruptcy filings, marriage records, and inheritance disputes.

The pipeline's core algorithm was not a predictive model. It was an optimization engine that matched welfare beneficiaries to specific antique lots based on a psychological profile derived from their consumption patterns, medical histories, and correspondence with caseworkers. The affinity score was not a measure of financial need—it was a measure of emotional vulnerability. A beneficiary who had written a letter about missing a lost family heirloom would get a high score for a similar-looking vase. A beneficiary whose father had been a pottery merchant would get a high score for a Joseon-era bowl. A beneficiary who had expressed anger at the government's welfare cuts would get a high score for a piece that could be resold at a premium, creating a sudden windfall that might justify a reduction in benefits.

Natsuki stared at the screen. Her reflection stared back—a pale, hollow-eyed woman in a beige cardigan, the kind of woman who had spent two years trying to outsmart a machine that had already outsmarted her.

The pipeline's last logged transaction was dated August 17, 10:47 PM—the night before Kuroda died. The care unit was #2027-0583, a 68-year-old woman named Park Sun-hee, a resident of Fukuoka who had once been the matriarch of the Kim textile dynasty. Her affinity score was 0.91. The matched antique was a "rare Goryeo dynasty inlaid celadon incense burner," estimated value 280 million won. The auction was scheduled for September 2 in a private gallery in Gangnam.

The beneficiary's benefit adjustment had been approved that same afternoon—a 3.4% increase due to "medical expenses." The approval was signed by a caseworker whose name Natsuki recognized: Shimizu, the same woman who had signed her mother's reduction notice four years ago.

Natsuki leaned back in her chair. The air conditioner clicked off, and the room fell silent. Through the window, she could see the gray sprawl of Fukuoka's harbor, the container cranes standing like skeletal dinosaurs against the evening sky. Somewhere beyond that water was Busan, and somewhere beyond Busan was a secret that had killed her supervisor and was about to consume a dead woman's memory.

She copied the Chrysanthemum data onto a USB drive—encrypted, with a key she generated offline. She deleted her browsing history. She locked the drawer again, with Kuroda's notebook still inside.

Then she walked to the elevator, pressed the ground-floor button, and waited. The doors opened on a young man in a delivery uniform, holding a cardboard box. He smiled at her. She smiled back. She stepped into the elevator, and as the doors closed, she caught a glimpse of his badge—a temporary ID with the name "Yi" and a logo she had never seen before.

The elevator descended. Natsuki's phone buzzed in her pocket. A text from an unknown number, with no sender ID:

"You found the kiln. Now find the fire."

The doors opened on the lobby. The fluorescent lights flickered once, twice, and then held steady. Natsuki walked out into the August heat, the USB drive burning in her palm like a fragment of a shattered bowl, and she knew, with a certainty that had nothing to do with data, that she was no longer a validator of anything.

She was a variable in someone else's algorithm.

And that algorithm was just getting started.