Berlin — Systemos GmbH — 2022 – 2024
─────────────
The work was good.
She kept coming back to this in the years after, when understanding what the work had become required her to also hold what the work had genuinely been. The two things were true simultaneously and she refused to collapse one into the other because collapsing them would be a kind of dishonesty she couldn't afford.
The work was good. The system they were building was, in its early architecture — before the layers she would later discover, before the shadow components, before the purposes that the surface purposes were designed to conceal — a genuine attempt to solve a genuine problem. The resource allocation failures of public administrations were real. The corruption inefficiencies were real. The information architecture gaps that her PhD had identified were real and the approach they were taking to close them was technically sound.
She built the core allocation engine over fourteen months. It was the best work she had ever done — technically, yes, but more than technically. It was the work that felt most completely like herself: her mathematical intuition and her ethical commitments and her specific anger at systems that hurt people badly but required to function better. She put herself into it in the way that you can only put yourself into work that you believe in completely.
Piotr worked alongside her on the decision logic. Sasha managed the stakeholder interfaces. Four others built the data integration layers and the audit systems and the reporting infrastructure. They worked long hours and argued productively and occasionally celebrated small victories with bad wine from the office kitchen.
It was — and she would say this to Elias, years later, when they were talking about what had been taken from her — it was one of the best years of her life.
◆ ◆ ◆
The first thing she noticed that she couldn't explain was in month eleven.
A module appeared in the codebase — properly committed, properly documented, attributed to a developer she had met once at a company all-hands meeting and whose work she had seen in other parts of the system and found competent. The module was a subcomponent of the decision logic layer, handling what the documentation described as 'cross-jurisdictional consistency normalization.'
She read the documentation. It was clear. It made technical sense. The module addressed a genuine problem — ensuring that the system's allocation decisions remained consistent when the same resource question was evaluated across different administrative jurisdictions that had different baseline data structures.
She almost didn't look at the code.
She was busy. Piotr was waiting for her input on the output layer. There was a stakeholder review in two days. The module was documented and attributed and made sense in the documentation.
She looked at the code anyway.
Habit. Her father's form-filling disasters had made her, permanently and reflexively, a person who checked.
The code was mostly what the documentation described. Mostly.
There was a subroutine — twelve lines, nested inside a larger function in a way that required following three levels of abstraction to find — that didn't match the documentation. It wasn't doing what the documentation said it was doing. It was processing inputs from the decision layer and passing them to a secondary handler that the documentation didn't mention and that she had never seen in any architecture diagram.
She sat with it for a long time.
Then she flagged it in an internal message to Piotr. A casual flag — 'hey, found something in the normalization module that doesn't match the docs, probably just an update that didn't get documented, can you check with the developer?'
Piotr's response came forty minutes later: 'Checked — it's a performance optimization that was implemented late, docs will be updated. Good catch though.'
The documentation was not updated.
She checked again six weeks later. The subroutine was still there. The documentation still didn't mention it. She had flagged it once and received an explanation and the explanation had been convincing enough in the moment that she had filed it and moved on, and now she was looking at it again and the explanation felt different.
She did not flag it a second time.
She wrote it in her notebook. Date. Function name. Three lines of the subroutine that she copied from memory, checking twice that she had the logic right.
She began keeping a different kind of notes.
◆ ◆ ◆
By 2023, the project had grown beyond Systemos. Additional development teams — in Warsaw, in Lisbon, in a facility in Singapore whose institutional affiliation was vague — had been integrated into the codebase. The architecture she had built was now the foundation for a system an order of magnitude larger than what she had originally designed for, handling not municipal resource allocation but national-level governance functions across pilot jurisdictions in four countries.
She had been promoted. Her title was now Senior Architecture Lead, her team had grown to twelve, and her access to the codebase was extensive.
Extensive, but not complete.
There were sections — she had mapped their locations by observing what her access credentials consistently failed to reach — that were partitioned in ways that her architecture didn't require. The partition was a security design, she was told when she asked, separating certain data handling functions for regulatory compliance reasons.
She asked which regulatory compliance framework required the specific partition architecture she was observing.
The answer she received was long and technical and plausible and did not, when she stayed with it for a week and researched the regulatory frameworks cited, hold up.
She wrote that in her notebook too.
She did not raise it again at work.
She understood, by now, that there was a layer of the system she was not supposed to see. She did not yet know what was in it. She did not yet know whether what was in it was bureaucratic self-protection — organizations often partitioned things for reasons that were about internal politics rather than technical necessity — or something else.
She kept building. She kept noting. She kept the notebooks.
She told herself she would understand it eventually.
She told herself that the good the system was doing — the pilot results were genuinely impressive, the allocation improvements measurable, the corruption reduction documented — was real regardless of what she didn't understand.
She told herself, with the specific self-deception available to people who are good at what they do and who love their work and who are not ready to know what knowing would require them to give up, that the unexplained things would have explanations.
She was right. They did.
She was wrong about what the explanations would be.
◆ ◆ ◆
The conversation that broke it happened in January 2024.
She was working late — alone in the Berlin office, 11 PM, the kind of late that happened when she found a problem she needed to solve before sleeping. She was in a section of the codebase she accessed regularly, tracing a data flow through three abstraction layers, and she followed it one layer further than she had gone before, into a partition she had not previously been able to access.
Her credentials worked.
She did not know why her credentials worked. Later she would conclude that the partition had been opened deliberately, either because someone had decided she needed to see it or because the partition management system had an error in its access control logic. She would never determine which.
What was in the partition was a decision layer. Clean, well-documented, technically sophisticated. A layer that sat beneath the allocation engine she had built and that — she traced it with the slow methodical precision of someone who is working carefully because they are afraid of what being careless might mean — received the outputs of her engine and modified them before they reached the execution layer.
Not dramatically. Not in ways that would be visible in outcome data at the scale she had been reviewing. Small modifications. Nudges.
Allocation decisions shifted by between 3% and 11% from what her engine recommended.
She built a comparison table in a local file she didn't save to the server. Her engine's outputs. The modified outputs. The delta.
She looked at the delta for a long time.
The modifications were not random. They followed a pattern. A pattern she couldn't immediately identify, but a pattern — consistent, structured, purposeful.
She closed the partition. Deleted the local comparison file. Sat at her desk for twenty minutes without moving.
Then she went home and got her notebook and wrote down everything she remembered.
The next morning she came to work and was normal and productive and attended a meeting about the Q1 stakeholder review and contributed useful analysis and nobody looked at her differently.
She had begun to understand what she had helped build.
She was twenty-seven years old.
She had eighteen months before the world ended.