He filed a sick day.
It was the first sick day Marcus had taken in fourteen months at Helix, which meant Gordon Fitch expressed concern in the slightly alarmed way that managers expressed concern when a quiet reliable employee showed any sign of being a human with a human body. Marcus texted back a simple *stomach thing, back tomorrow* and closed the conversation.
Then he cleared his desk, made a fresh pot of coffee, and sat down to work.
He had a target.
For three months, his crawler had been collecting public procurement data, and in that data there was a specific anomaly he had flagged and then set aside because his old understanding of it was insufficient to fully interpret it. A cluster of government IT contracts in three adjacent states, awarded to a rotating family of small vendors over eight years, with a pattern in the award timing that didn't match competitive bidding cycles. It was the kind of thing that a competent forensic accountant could eventually piece together with enough time. Marcus had never had the tools to move beyond the pattern-recognition stage.
He had the tools now.
He rebuilt the indexing layer again — not because the previous version was bad, but because the previous version was a prototype and what he was building today was the real thing. He wrote the new indexing engine in two hours. It was half the size of the previous version and substantially more capable. He looked at it when he was done and had the same vertiginous feeling as the night before: *I made this?* Then the feeling passed and he moved on.
By noon he had a cross-referencing module that could pull vendor registration data, state contractor filings, and public campaign finance records, normalize them against each other, and generate a network graph of relationships. By 2 PM the graph was running. By 3:30 he was looking at results.
The anomaly was not eight years old. It was eighteen.
The vendor rotation was not a coincidence or a quirk of procurement bureaucracy. It was a coordinated structure — a wheel with several spokes — in which approximately $340 million in public IT contracts had been cycled through a network of nominally independent companies, all of which traced back, through two or three layers of incorporation, to four individuals. Marcus found their names without difficulty. Two of them were on LinkedIn. One had a modest web presence as a "procurement consultant." The fourth had no public web presence at all.
Marcus looked at the graph for a long time.
This was not, technically, his problem. He was a contract programmer at a mid-sized data company, and what he was looking at was evidence of long-running public fraud that was properly the domain of investigative journalists or federal investigators, neither of which he was.
He also had thirty-one thousand dollars of reasons to care about finding sources of income that weren't his current contract rate.
He thought carefully. He was not reckless by nature — not impulsive, not aggressive. He was deliberate. He thought about what he was looking at, what it was worth, who it was worth it to, and what the risks of each possible action were.
He did not think about going to the four individuals. That was the kind of thinking that ended careers and occasionally ended people.
He thought about journalists. There were investigative outlets that paid for documented tips — not enormous sums, but something. More importantly, he thought about the data itself as a template. If this pattern existed in three states, it almost certainly existed elsewhere. The methodology he had built this morning — the cross-referencing engine, the normalization layer, the network graph generator — was generalizable. It could be applied to any procurement data. Any state. Any sector.
He was not looking at a one-time finding.
He was looking at a product.
He opened a new document and started writing.
---
He called it **Threadline**. The name came to him while he was writing the spec and he liked it well enough to keep it — it had the right combination of approachability and technical implication.
The concept was straightforward: a platform that ingested public financial data from government sources and surfaced structural anomalies. Not a journalism tool, not an investigation firm. Something between a research utility and a monitoring service. He could see exactly two customer segments immediately: investigative journalists, and compliance officers at companies that did business with governments and needed to understand the competitive landscape they were operating in.
Neither segment was enormous. Neither segment was irrelevant. He had run the procurement database long enough to know that the world was extremely full of $340 million anomalies that nobody had the tools or time to find.
By 7 PM he had a detailed product spec, a rough financial model, and the architectural design of the first version. He looked at the architectural design and recognized immediately that the previous version of himself — last Tuesday's version — would have needed six weeks to build it. He thought the current version might need eight days.
He checked the System display. It had been visible at the edge of his vision intermittently throughout the day, a passive presence, occasionally updating with something brief.
Now it showed:
---
**PERFORMANCE LOG — DAY 2**
**Architecture Authority** — Lv. 1 → **Lv. 2** *(design complexity threshold exceeded)*
**Inference Engine** — Lv. 3 *(stable)*
**Simulation Depth** — Lv. 1 → **Lv. 2** *(multi-system modeling confirmed)*
**FIRST GATE: 61% complete.**
Remaining condition: *external validation of output quality differential.*
---
Marcus read the last line.
*External validation of output quality differential.*
In other words: someone else had to recognize that what he had built was beyond normal.
He looked at his product spec. He thought about the right person to show it to.
He had exactly one contact in the startup world worth anything — a woman named Sera Okafor who had left a VC firm two years ago to run a small tech accelerator focused on civic technology. He had met her at a conference, had one follow-up conversation, and they had stayed loosely in touch in the way that professional acquaintances did, occasional likes on LinkedIn posts, a brief exchange when something relevant came across his feed.
He wrote her an email. He spent forty minutes on it. He described what he had found and what he had built and what he thought it could become, in language that was precise and direct and contained no hedging. He attached a three-page summary doc and the architecture diagram.
He sent it at 8:47 PM.
He went to bed before midnight for the first time in three days.