The second Northern Virginia meeting had a different energy.
The same room, the same frosted glass door. But Pemberton had brought someone new: a woman who was introduced only as Dr. Chen, who said nothing for the first thirty minutes and whose stillness had the quality of a very good engineer reading an unfamiliar codebase — absorbing everything before concluding anything.
Pemberton opened with the provision Marcus had requested.
"We've drafted language," he said, pushing a single page across the table. "Read it before I explain it."
Marcus read it.
The provision was narrow and precise. It allowed Marcus to raise a substantive ethical objection to a specific work thread, which would trigger a thirty-day review period during which the thread would be paused. At the end of the review, a three-person panel would make a determination. One panel member would be from the working group, one would be from an external oversight office, and one would be Marcus.
He read it twice.
"The panel composition," he said. "The external oversight member — what office?"
"Congressional oversight staff," Pemberton said. "Not a political appointment. Career staff."
"And their determination is binding?"
"Majority binding. If you and the oversight member are aligned, the thread stops."
Marcus looked at the page. This was more than he had expected. He had expected a provision that created a process while ensuring the outcome was always the same. This provision created a process in which his vote genuinely mattered.
Which meant they wanted him cooperative enough to actually use the provision honestly, rather than hostile enough to find other ways to obstruct.
He filed this inference carefully.
"Acceptable," he said. "My attorney will want to review the final language."
"Of course." Pemberton moved on. "The scope. We discussed a unified integration layer for classified data sources. I want to expand that slightly." He turned to Dr. Chen.
Dr. Chen spoke for the first time. Her voice was precise and unhurried, and it was immediately clear that she was the technical mind in the room in the same way that Marcus was the technical mind in his own company.
"The current state," she said, "is that we have three separate analytical platforms, each handling one category of data, each producing outputs that require manual reconciliation by analysts. The manual reconciliation step is the bottleneck. We have forty-three analysts doing reconciliation work that should be automated. What we're missing is a unified data model that lets the three platforms speak to each other."
"What are the three platforms?" Marcus said.
"Financial flow tracking, corporate registry analysis, and what we call signals-adjacent — which is commercial data combined with certain intercepted communications that have financial relevance."
"Signals-adjacent meaning you have access to intercepted communications?"
"Relevant fragments only, properly handled under FISA and relevant oversight." She said it without defensiveness — as a technical fact. "For our purposes, they're structured data inputs. The question is how to normalize them against the financial and registry data."
Marcus thought about this. "The normalization problem for signals-adjacent data is fundamentally different from the normalization problem for financial flows. The data models are incompatible at the schema level — you can't run a standard ETL process across them. You'd need an intermediate representation."
Dr. Chen looked at him. Something shifted in her expression — recognition, possibly.
"We've been calling it an ontology layer," she said. "We haven't been able to build one that's general enough to be useful."
"Because you've been trying to build it top-down," Marcus said. "You've been trying to design a unified schema and then force your data into it. That fails because the data types have incompatible semantics, not just incompatible formats."
"Yes," she said. "That's exactly right."
"You need to build it bottom-up. Start with the relationships you care about — entity X transferred value to entity Y, entity X controls entity Y, entity X appeared in context Z — and build a relationship graph where those propositions are the nodes rather than the entities themselves. Then your entities are just attributes of the relationships, and the schema incompatibility disappears."
Dr. Chen was very still.
"That's—" She stopped. Started again. "We've had three teams attempt to solve this problem over eighteen months. No one has suggested that approach."
"It's non-obvious if you're coming from a data engineering background," Marcus said. "It's more natural if you think of it as a graph traversal problem rather than a schema mapping problem."
Pemberton was watching Marcus with an expression he hadn't had in the first meeting. It was not admiration. It was something more like recalculation.
"Nine months," Marcus said. "I'll have a working prototype. Not a demonstration — a working system that your forty-three analysts can use."
Pemberton glanced at Dr. Chen. She nodded once.
"We have a deal," Pemberton said.
---
On the flight home, the System updated steadily:
---
**Architecture Authority Lv. 5 → Lv. 6** *(novel architectural framework proposed for domain-crossing problem)*
**Third Gate: 44% complete.**
*Note: current trajectory involves access to classified systems and personnel. Operator is advised that Real-World Integration events in this context carry amplified downstream consequence weighting.*
---
Marcus read the note twice. *Amplified downstream consequence weighting.*
He understood. He was no longer building tools that surfaced what had already happened. He was building tools that would affect what happened next.
He put the laptop away and looked out the window at the dark Atlantic below.