The phone buzzed against the nightstand at 6:14 in the morning and Mira Alvi opened her eyes in the dark.

Her bedroom faced east, toward Flatbush Avenue, and at this hour in March there was nothing at the window. No light, no color. Just the flat black rectangle of glass and, beyond it, the amber sodium glow of streetlamps two stories below. The radiator ticked in the corner, working through its ancient logic of knocks and pauses. She had lived in this apartment for three years and still could not predict when the heat would come on.

Four hours of sleep. She'd filed a story at two, a long investigation into wash trading on three minor crypto exchanges that her editor would sit on for weeks. Before that, a bowl of pho from the place on Vanderbilt that stayed open past midnight, eaten standing at the kitchen counter while she reread her own copy. Before that, five hours of screen time, interviews transcribed, sources triple-checked, the particular exhaustion of precision. The phone buzzed again. She reached for it.

The text was from a number she hadn't seen in eight months. Ray Campos, a compliance analyst she'd cultivated during the Iran bet story. He'd gone quiet after the Senate hearings. People did that. Sources burned out or got scared or simply decided they'd given enough. She'd stopped reaching out in October.

The message read: "Check Parallax. Now."

No greeting. No context. She sat up.

The apartment was cold. She reached for the sweatshirt draped over the bedpost and opened the Parallax app before her feet hit the floor. Professional habit. She navigated past the landing page, past the featured indices, past the usual noise of political contracts and Fed rate bets.

She found PX-8891 in the trending tab. Top of the list.

"Significant military action by a state actor against a sovereign nation initiated by 23:59:59 UTC Friday."

Trading at $0.22. Up from $0.03 at creation. Volume had already crossed $11 million.

She read the contract language again. State actor. Sovereign nation. Significant military action. The phrasing was LUMEN's, Parallax's contract-generation AI. No human had written those words. The system had assembled them from whatever signals it was pulling, decided the probability of a major military strike was worth pricing, and published the contract to the world. Automatically. At 2:14 in the morning.

Mira scrolled to the market detail panel and found the number that stopped her.

Agent share: 62%.

Sixty-two percent of the volume on this contract was coming from autonomous trading agents. AI systems with their own wallets and trading strategies, plugged into Parallax through its agent marketplace API. They'd been active since minutes after the contract went live.

She sat on the edge of her bed.

Mira had first encountered prediction markets on election night, 2024. She'd been twenty-six, a second-year reporter at a fintech trade publication, assigned to cover the regulatory angle of crypto betting platforms. Nobody at the outlet cared much about Polymarket. It was a small beat, almost filler. But she'd been watching the platform's election markets that night from her desk in a WeWork on Fulton Street, and she'd seen the prices move before anything else. Before the AP. Before the networks. Before the exit polls coalesced into a narrative. At 7:42 PM Eastern, Polymarket's Trump contract spiked from $0.52 to $0.61 on no visible news. She refreshed CNN, NBC, the Times. Nothing. The prices kept climbing. By 8:15, the contract was at $0.74. The networks were still calling it too early to tell. She remembered the feeling: the slow, disorienting realization that a gambling website staffed by anonymous speculators was processing information faster than every newsroom in the country.

She'd written a piece about it the next morning. Her editor cut it by half. Nobody at the publication thought prediction markets were a real story yet. She quit the fintech outlet four months later and spent six weeks freelancing before The Atlantic picked up her first pitch: a reported essay on how offshore prediction markets were becoming a shadow information system, pricing events that traditional media couldn't or wouldn't cover. The piece ran at 3,400 words and paid enough to cover two months of rent. More important, it introduced her to the world she would spend the next two years reporting on.

By 2025, prediction markets had gone from curiosity to infrastructure, though the early platforms had clear ceilings. Kalshi was listing hundreds of contracts on economic indicators, weather events, regulatory decisions, but it operated under CFTC constraints that limited what it could price. Polymarket, despite operating in a legal gray zone, was processing billions in annual volume, but it was centralized, fragile, and one subpoena away from collapse. A cottage industry of analysts, traders, and academics had formed around the platforms, debating calibration methods and market efficiency on Twitter and Substack. By 2027, the cottage industry had professionalized. Hedge funds subscribed to market-derived sentiment indices. Reinsurance companies priced catastrophe bonds off prediction market signals. Intelligence agencies monitored contract movements as open-source indicators. Three universities had launched prediction market research centers. Mira knew most of the serious people in the space by name. She'd interviewed them, sourced from them, and occasionally argued with them at conferences where the panels had titles like "Price Discovery or Reality Production?" and nobody could agree on the answer.

In late February 2026, in the hours before the United States and Israel launched a second strike on Iran, somebody started buying YES shares on Polymarket's Iran strike contract. More than 150 accounts. Individual bets over a thousand dollars each. Sixteen wallets that each cleared more than a hundred thousand. The buying started before any public indication of the operation, before any leak, before any journalist had a whisper.

Amy Fan at the Times broke that story. Her front-page investigation traced the wallet addresses, mapped the timing, established that someone with advance knowledge of a military strike had used a prediction market to profit from it. A blockchain forensics firm in Singapore clustered the wallets by funding source. The circumstantial case was airtight. The direct proof was, as always, just out of reach.

Senator Caldwell called it "worse than insider trading" in an interview. The CFTC opened an inquiry. Mira had been working the same angle, three weeks in, when the Times piece dropped. She killed her draft and pivoted to the regulatory fallout, a 6,000-word piece for The Atlantic on the institutional failure to anticipate what prediction markets had become. It was good work. It was not the story she'd wanted to write.

There was a particular quality to the Iran reporting that she still thought about. The moral vertigo of it. She'd sat at this same kitchen table on a Friday afternoon, watching money flow in real time toward a war. Hundreds of thousands of dollars piling into YES contracts while, somewhere in the eastern Mediterranean, pilots were presumably running final briefings. She remembered refreshing the Polymarket page every thirty seconds, watching the price climb through the afternoon, each cent representing another cluster of bets by people who seemed to know what was coming. She'd called a source at the Pentagon press office. The source said there was nothing to report. The next morning, the first satellite images of the strike appeared on social media.

The market knew. The money knew. And the knowledge was just sitting there, priced at sixty-seven cents, visible to anyone with an internet connection, and no one could do anything with it because the knowledge was expressed as a bet, and a bet was legal. She'd tried to explain this to her mother on the phone that weekend. Her mother, a retired pediatrician in Dearborn, had listened patiently and then asked, "So it's like gambling on whether someone gets sick?" Mira had said, "It's like gambling on whether someone gets shot, and the gamblers know about the gun before the victim does." Her mother had gone quiet.

The Iran story was simple, in retrospect. People with information bet on what they knew. Prediction markets as an insider trading venue. The analogy to stock markets was direct enough that regulators could grasp it.

A year later, in early 2027, she'd reported a subtler version of the same problem. ViralEdge Therapeutics, a Lagos-based biotech, had been running late-stage trials on a malaria vaccine. A cluster of Parallax contracts appeared pricing the likelihood of WHO prequalification. The contracts said the vaccine would fail. A European hedge fund saw the signal, shorted the stock of ViralEdge's manufacturing partner in Frankfurt, and the price collapsed. The vaccine ultimately passed prequalification, with conditions, but by then the stock damage was done and the manufacturing partner pulled out of the deal. The terms were restructured at a fraction of their original value. Three distribution agreements in West Africa fell through. The vaccine still existed, but the infrastructure to deliver it had been crippled by a price movement on a screen six thousand miles away.

The question Mira could never answer was whether anyone had intended this. Had someone created the Parallax contracts knowing a hedge fund would act on the signal? Or had the contracts simply reflected a legitimate doubt about the trials, and the cascade was an unintended consequence of markets doing what markets do? Either way, the result was the same. Mira had spent three weeks in Abuja tracing the contracts back to their funding sources. The piece ran at 4,200 words. The hedge fund declined to comment. The WHO issued a statement expressing concern about "speculative interference in global health procurement." Nobody was charged with anything, because nobody could agree on what law had been broken.

The Abuja story was harder to explain than Iran. There was no gunshot, no military strike, no satellite imagery. Just prices on a screen, capital flowing through prediction markets into equity markets and back again, and a vaccine program that survived technically but was damaged in ways that took months to quantify.

What Mira was looking at now was something else.

She walked to the kitchen and started the electric kettle. The apartment in Prospect Heights was a one-bedroom railroad flat, narrow and deep, the kind of place where you could stand at the stove and see the front door. She lived alone. Had for two years, since the lease with Elena expired and neither of them suggested renewing. The kitchen counter served as her second desk, cluttered with a charging cable, a stack of FOIA printouts she hadn't filed, and a half-empty bag of ground coffee from a roaster on Bergen Street.

Pre-dawn light was starting at the edges of the window, the sky shifting from black to the dark blue-gray that March mornings in Brooklyn produced, a color that always seemed to promise nothing. She opened her laptop on the counter.

She logged into Parallax on the desktop, where the data was denser. The interface was clean, almost austere: dark background, white text, the kind of design language that signaled seriousness. Elias Marino had reportedly insisted on the aesthetic during development. No gamification, no confetti animations when trades cleared, no dopamine tricks. Parallax positioned itself as an information market, a tool for collective intelligence, and it looked the part. The first time Mira had seen the platform, during a demo at a crypto conference in 2027, she'd thought it looked like a Bloomberg terminal designed by someone who'd read too much Habermas.

The order book for PX-8891 filled the screen. She could see the buy wall at $0.21, the ask thinning above $0.23. The chart showed the spike from creation: four hours of vertical movement, then a consolidation, then another push up. Classic momentum pattern. But the composition was wrong.

On Polymarket, back in 2024 and even in 2026, the actors were human. Whales and retail bettors and a few early algorithmic traders working with basic scripts. The Iran wallets had been operated by people. You could see it in the timing. Irregular intervals between orders. Clusters around round numbers. The behavioral fingerprints of someone sitting at a keyboard.

Parallax was different. Parallax had launched its agent marketplace in late 2027, and the platform had transformed. The idea, as Elias Marino articulated it in a widely circulated blog post, was that autonomous AI agents represented the next evolution of market participation. They could process information faster, trade more rationally, and maintain positions without the emotional biases that plagued human traders. The marketplace was open: any developer could register an agent, connect it to the Parallax API, fund its wallet, and let it trade. No human needed to approve individual orders. No human needed to monitor performance.

Within six months, more than 3,000 agents had registered. Some were simple arbitrage bots, buying contracts on Parallax and selling equivalent positions on Kalshi or smaller offshore platforms to capture price discrepancies. Some were sophisticated multi-signal systems backed by venture capital, running proprietary models on dedicated GPU clusters. Some were experimental research projects from university labs. A team at ETH Zurich had built an agent that traded exclusively on climate-related contracts using atmospheric modeling data. A startup in Tel Aviv ran a cluster of agents focused on Middle Eastern geopolitical events, fed by a proprietary Arabic-language news analysis pipeline. And they all had access to the same contracts, the same order books, the same data.

The regulatory landscape was incoherent. Kalshi had spent three years in federal court for the right to list political event contracts. Polymarket, offshore, ignored regulation entirely. And Parallax, with its agent marketplace, had opened a third front. A former CFTC commissioner Mira interviewed had laughed: "We spent four years deciding if an election contract was a swap or a binary option. You're asking me about robot traders on a decentralized platform incorporated in Singapore. We don't even have a category for that."

Sixty-two percent agent share on a contract about a military strike.

Mira called her editor.

David Auerbach answered on the fourth ring. He sounded like a man who had been asleep ten minutes ago, which he had.

"It's six-twenty on a Sunday, Mira."

"I need to talk through something."

"Is someone dead?"

"Maybe. Not yet." She paused. "Do you know what Parallax is?"

"Prediction market. The AI one. Elias Marino's company. We ran a profile on him in September."

"Right. There's a contract on Parallax right now predicting a major military strike. A state actor against a sovereign nation. It was generated four hours ago by their AI, and it's already at twenty-two cents with eleven million in volume."

Silence on the line.

She could hear him sitting up, the creak of bedsprings. David lived in Park Slope with his wife and two daughters. He'd been editing investigative features at The Atlantic for eleven years and had the particular patience of a man who had learned that the best stories took the longest to prove.

"So twenty-two cents means the market thinks there's a twenty-two percent chance of a military strike," David said. "That's how these contracts work, right? Dollar payout if the event occurs, zero if it doesn't."

"Right. And it went from three cents to twenty-two in four hours."

"What's driving it?"

"That's the problem. Sixty-two percent of the volume is coming from autonomous agents. AI trading systems. Parallax lets them trade on the platform directly, no human approval required."

David was quiet for a moment. "So sixty-two percent of the money is from algorithms. What's the other thirty-eight percent? Retail?"

"Mix of retail and institutional. Some whale accounts. But the agents moved first. They were on this contract within minutes of creation."

"Okay, so here's my question." David's voice had shifted. He was awake now, fully, the editor replacing the man who'd been sleeping. "Why is this different from any other algorithmic trading? High-frequency traders move faster than humans in equity markets. That's been true for twenty years. Renaissance Technologies makes billions by processing information faster than everyone else. Why does it matter that bots are trading on a prediction about a war?"

It was the right question. David always asked the right question. He had a talent for finding the assumption buried inside the claim and pulling it out for examination. It was what made him a good editor and an occasionally infuriating one.

Mira had been organizing the answer since she'd first seen the number.

"In 2026, when the Iran bets happened, we could trace it back to people. Wallets controlled by people who had information. The theory was that someone connected to the operation, or someone who'd been briefed, leaked it. People made bets. We followed the money to the source."

"And now it's bots making the bets."

"Agents. Not scripts running predefined rules. These are autonomous systems that form their own analysis, decide what to trade, and execute without human approval. And the question is whether those agents are picking up on something real, something in the data that a human analyst would also see. Satellite imagery, diplomatic chatter, social media patterns. Or whether something else is happening."

"Something else meaning what?"

Mira looked at the order book on her screen. A new cluster of buys had appeared at $0.22, pushing the ask up. She watched an agent identified as SIGMA-4 place a $28,000 order.

"Meaning someone could be feeding signals to the agents. Or to LUMEN directly. The AI that generates the contracts. If you could influence what contracts LUMEN creates, you could front-run your own prediction."

"You're saying someone could trick the AI into creating a prediction about a war, then bet on the prediction they caused."

"Potentially."

"And then what? The war doesn't happen, the contract settles at zero, and the manipulator loses their money?"

"Unless they've structured their positions to profit on the volatility itself. Or unless the prediction creates real-world effects. Insurance companies adjust coverage based on prediction market signals. Defense contractors watch these platforms. If a twenty-two-cent contract on a military strike reaches forty or fifty cents, that changes behavior in the physical world."

"Give me an example."

"Last year, a Parallax stability index for the Red Sea shipping corridor moved four points in one afternoon. A Nordic insurance consortium adjusted their coverage automatically based on the signal. Three shipping companies lost their insurance. They had to reroute. It disrupted logistics in East Africa for almost two weeks."

"And was that deliberate?"

"I don't know. That's the point. The system is complex enough that effects can cascade without anyone intending them. Or someone could intend them and the complexity would provide cover."

David was quiet for a long time. She could hear him thinking, which with David sounded like breathing with longer pauses between the breaths.

"Let me push back for a second," he said. "You're looking at a prediction market that says there's a twenty-two percent chance of a military strike. That's not a high probability. That means the market thinks there's a seventy-eight percent chance nothing happens. Is it possible this is just the market doing what it's supposed to do? Processing information, arriving at a probability, and the probability happens to be scary because the subject matter is scary?"

"It's possible. That's why I'm calling you at six in the morning instead of writing a story. I don't have enough yet. But the agent concentration bothers me. And the speed. Three minutes from contract creation to the first trade. That's not normal."

"What's normal?"

"Ten to fifteen minutes. Sometimes longer. And most LUMEN-generated contracts never attract significant volume at all."

"So something was ready for this one."

"Something was waiting for it."

"That sounds like the Iran story again. But worse."

"The Iran story was a ten on the complexity scale. This might be a hundred."

"What do you need?"

"Time. And I need to make some calls. I was working on a Parallax feature six months ago. I shelved it because I couldn't get anyone inside the company to talk on record. I have sources who might talk off record now."

"How soon can you have something?"

"I don't know yet. Let me work today and I'll call you tonight."

"Mira." He paused. "If this is what you think it is, we need to be very careful about what we publish and when. A story about market manipulation is one thing. A story that itself moves a market about a war is something else."

"I know."

She hung up and poured the water over her coffee grounds. French press, the only method she trusted. She let it steep while she opened her contacts and scrolled.

Hana Javed. Former chief scientist at Parallax, the person who had designed LUMEN's core architecture. She'd interviewed Hana once, back when she first started reporting the Parallax story, before she'd shelved it. They'd spoken for forty-five minutes over a Zoom call, Hana in London, Mira in this apartment. Hana had been cautious, speaking in generalities about AI-generated prediction contracts, never quite criticizing the platform she'd helped build. She'd deflected direct questions about her departure with the kind of measured non-answers that told Mira more than the words themselves. "I left because I wanted to pursue other work." "My time at Parallax was valuable and I wish them well." Corporate language from a woman who clearly did not think in corporate language.

But she'd left Parallax. People who leave voluntarily usually have reasons, and those reasons sometimes become stories. Mira had kept Hana's number and checked in twice over the following months, brief emails that Hana answered politely without revealing anything new. The relationship was dormant. It was time to wake it up.

Mira called. The line rang six times and went to voicemail. London number. It was past eleven there. Sunday morning. She left a message.

"Hana, it's Mira Alvi. I know it's been a while. I need to talk to you about something on Parallax. Contract PX-8891. Call me when you get this."

She pressed the plunger on the French press and poured. The coffee was bitter and too strong and exactly what she needed. She drank it standing at the counter, looking out the window at the slowly brightening street. A man walked past with a dog, both of them hunched against the cold. The streetlamp outside her building switched off, the timer deciding that enough light had arrived. It hadn't, really. The sky was still the color of old steel.

She thought about the Times. They had resources she didn't: a newsroom, a legal team, institutional relationships with the intelligence community. But the Times was also competition. If this turned into a story, they'd be racing. She decided to work the leads first.

Back at the laptop, she started building a timeline. This was how she worked: sequence first, theory second. What happened, then what happened, then what happened. She opened a new document and typed:

02:14 UTC: LUMEN generates PX-8891. No human review. 02:17 UTC: First trade. Three minutes and forty-one seconds after creation. 02:14-03:00 UTC: Wave of agent buy orders. Price: $0.03 to $0.12. 03:00-06:00 UTC: Continued buying, mix of agent and human. Price: $0.12 to $0.22. Volume at 06:14 UTC: $11.3M. Agent share: 62%.

She studied the timeline. Three minutes and forty-one seconds from contract creation to the first trade. That was fast. Most LUMEN-generated contracts sat for ten or fifteen minutes before anyone noticed them. Some never attracted volume at all.

Three minutes and forty-one seconds meant someone, or something, was watching for exactly this type of contract. Waiting for it.

She thought about the Maduro bet. That one had been messier. On January 3, 2026, hours before U.S. troops captured Nicolás Maduro, someone had placed a large position on a Polymarket contract pricing his removal from power. The tip had come to her from a Venezuelan journalist named Carla Soto, who'd been tracking exile-community betting patterns from Miami. Carla had noticed the position, a single wallet that put $32,000 into a contract that had been trading at pennies, and called Mira because she'd read her Polymarket election coverage. Mira spent two weeks verifying it, pulling transaction records, cross-referencing the wallet's funding chain with wire transfer databases. The wallet turned into nearly half a million dollars when the capture operation succeeded. The timing was suspect. The profit was obscene. But the regulatory apparatus still hadn't figured out how to treat prediction market bets as securities, so the CFTC referral had gone nowhere. The story ran as a 2,000-word feature. It was cited in three academic papers and changed nothing.

PX-8891 was different from all of these. The Maduro bet, the Iran wallets, the Abuja pharma contracts, the Bucharest election cascade in November where agents had bought far-right coalition victory contracts at 64% agent share and media outlets reported the price as a prediction and the coalition won by 2.3 points and nobody could say whether the market had predicted the outcome or produced it. Even the Polymarket election markets. Those were markets where humans created the contracts and humans placed the bets. You could follow the money to a person. You could subpoena records, interview witnesses, build a chain of evidence that ended at a human being who made a decision to place a trade. The investigative tools were old ones, refined by decades of financial journalism and securities enforcement.

PX-8891 was a contract created by an AI, traded predominantly by AIs, predicting an act of war. Where did you even begin an investigation? Who was the subject? The agents were software. Their developers might be traceable, or they might be anonymous, hidden behind shell companies and jurisdictional arbitrage. LUMEN itself was a system, not a person. It had generated the contract based on signals it had processed through layers of machine learning models. There was no memo to subpoena, no phone call to record, no source to meet in a parking garage. The entire chain, from signal to contract to trade, was mediated by software. And the price was still climbing.

She refreshed the page. $0.23. Volume: $11.8 million.

Mira picked up her phone and texted Ray Campos back.

"I see it. What am I looking at?"

The reply came in forty seconds. "Look at the agent cluster between 02:17 and 02:20. Thirty-one unique agents entered positions within three minutes of generation. Cross-reference with their trading history on other military contracts."

She pulled up the Parallax analytics dashboard, the public version that showed aggregate data without identifying individual agents by their full wallet addresses. She ran her OSINT crawler on the contract ID. The crawler was a custom tool she'd built with a developer friend in 2027, after the Abuja story: it scraped public blockchain data, cross-referenced wallet histories across platforms, flagged anomalous timing patterns, and output a structured summary that would have taken her three days to compile manually. Every investigative reporter she knew had some version of this now. The ones who didn't were already behind. She filtered for PX-8891, first hour, agent-only trades.

Ray was right. Thirty-one agents in the first three minutes. She clicked through to the agent profiles that were publicly visible. Most had trading histories going back months. Commodities contracts, geopolitical events, election outcomes. Standard algorithmic trading behavior. But she noticed something in the pattern. Fourteen of the thirty-one had placed positions on the same three contracts in the previous week: a Taiwan Strait shipping disruption contract, a NATO emergency session contract, and a contract on petroleum futures exceeding a specific price target.

Fourteen agents. The same three contracts. Then all fourteen among the first to pile into PX-8891.

Correlation wasn't causation. She knew that. She'd written those exact words in a dozen stories and meant them every time. Agents could share similar data sources, similar models, similar strategy architectures. They could independently arrive at the same conclusions from the same publicly available information. A spike in shipping disruption chatter, publicly visible satellite data showing naval movements, a trend in oil futures. Any sophisticated model might flag the same pattern. But fourteen out of thirty-one was a striking cluster, and the three prior contracts they'd traded formed a logical chain. Shipping disruptions. NATO response. Oil prices. Military action. Each contract a step in an escalation ladder. Each step feeding into the next.

She opened her notebook, the physical one, the spiral-bound reporter's pad she'd used since grad school. She wrote the lede she would never publish, the one that existed only to clarify her own thinking.

"An artificial intelligence system designed to predict the future has decided that a war is coming, and the robots believe it."

Too cute. She crossed it out. Tried again.

"On Sunday morning, a prediction market contract auto-generated by an AI and traded predominantly by autonomous agents placed the probability of a major military strike within five days at twenty-two percent and rising."

Better. Factual. Verifiable. Everything she knew. She stared at the words on the page and felt the familiar pull of the story taking shape, the moment when scattered observations begin to organize themselves into a structure. She'd learned to be suspicious of that feeling. The pattern that seems obvious at 6 AM sometimes dissolves by noon. The story that feels urgent on a Sunday morning sometimes turns out to be noise dressed up as signal. She'd killed a dozen stories in three years by forcing herself to wait, to verify, to resist the gravity of a good narrative.

But it wasn't the story yet. The story was who was behind the buying. Whether the agents were responding to real intelligence signals or manufactured ones. Whether LUMEN had generated this contract because the world was genuinely moving toward conflict or because someone had figured out how to make it generate what they wanted.

She finished her coffee and started a second pot. It was going to be a long Sunday. Through the kitchen window, the sky had lightened to the color of wet concrete. Across the street, the bodega's fluorescent sign flickered on. A garbage truck groaned down the block. The neighborhood was waking up, the Sunday rhythms: slower footsteps on the sidewalk, the distant clang of a church bell from St. Augustine's on Sterling Place, the smell of something baking from the apartment below hers. She should eat. She wouldn't, yet.

Her phone buzzed. Not Ray. Not Hana. A push notification from the Parallax app.

PX-8891: $0.24 ▲ | Vol: $12.1M

The number kept climbing. Somewhere in the world, the machines were betting on a war.