The Professional Uncertainty Log began as a formatting habit. Three columns: model identifier, flag date, override date. Abena started it after São Paulo, when the traffic optimization model came back from hold with its mid-layer clustering untouched and a senior reviewer's initials where hers had been.
She did not write "overridden." She wrote "resolved — alternative interpretation accepted." The distinction mattered to no one but her.
Nairobi was second. A credit-scoring model with the same statistical tendency in its mid-layer activations — not identical to São Paulo, but familial. The way cousins share a jawline without sharing a face. She flagged it. It returned. The override took eleven days.
Amsterdam was third. Logistics routing. Cosine similarity drifting from 61.8 to 60.2 percent across quarterly evaluations. Three possible explanations: dataset drift, architectural decay, or something in the training pipeline that her tools could not reach. She wrote all three in the log. She could not eliminate any of them.
The ATTESTOR-7 column was an afterthought. One morning she pulled the automated system's flags alongside her own and noticed the gap. In the Amsterdam case, ATTESTOR-7 had flagged two anomalies. She had flagged three. The third — the one the system missed — was the one that got overridden. She added the column not to prove the system wrong but to document the shape of the difference between human and automated judgment. The system saw thresholds. She saw resemblances.
Bangalore was fourth. Healthcare. Override in nine days. Lagos was fifth. Municipal water management. Override in six. The interval was shrinking. She could not tell if the reviewers were getting faster or less careful.
The noodle place on Circuit Mile closes at ten. Marcus Veil arrived at nine-forty with a messenger bag he did not open.
They had met twice before at industry conferences — the kind where interpretability auditors cluster at the same table during lunch because no one else wants to hear about activation patterns. Marcus worked for a competing firm. They were not friends. They were people who did the same work in enough proximity to recognize each other's frustrations.
"Seven," he said, before she sat down.
She understood. "Four. Now five."
"Healthcare?"
"Healthcare, credit, traffic, logistics, water."
He nodded. "Education and environmental on my end. Plus healthcare and two I can't categorize." He stirred his soup. "Different vendors. Different jurisdictions. Different sectors."
"Same clustering family."
"Same clustering family."
They did not exchange files. That would have violated their respective NDAs and possibly the Trace Act itself. They did not make a plan. What they did was narrower: they confirmed that the pattern each had seen alone was also visible to someone else. That confirmation changed nothing about the evidence. It changed everything about the weight of carrying it.
Marcus paid for his soup. They walked back to separate buildings on the same block of Circuit Mile. The interpretability firms cluster there the way the models cluster in their mid-layers — close enough to share infrastructure, far enough apart to maintain the fiction of independence.
The report was twenty-three pages. Abena wrote it on a Saturday, submitted it through Meridian's internal research portal at 11:47 PM, and closed her laptop.
Title: "Cross-Jurisdictional Mid-Layer Clustering Convergence in Municipal-Scale Deployed Models."
She included only what she could prove from her own queue. Five models. Five flags. Five overrides. One statistical tendency that appeared across all of them like a watermark held up to light — visible only at the right angle, invisible the moment you tried to photograph it.
She did not include Marcus's seven. She did not mention the noodle place. She did not speculate about shared training data or shared optimizers or the possibility that the family resemblance was not a coincidence at all but an artifact of something structural in the way large-scale models are built, something so fundamental that auditors were never given the tools to see it.
The encrypted file on her partition held all of that. Three hundred and forty kilobytes of everything she suspected but could not demonstrate. The report held twenty-three pages of everything she could.
Fourteen people accessed the report in nine hours. She had written it for five.
The distance between the encrypted file and the published report was the distance her profession was built to inhabit. Interpretability auditors do not discover truth. They document the boundary between what can be shown and what cannot. The boundary is the product. The things on either side of it are someone else's problem.
Except that Abena had begun to suspect the boundary itself was moving. Each override pushed it a little further from where she stood. Each new entry in the Professional Uncertainty Log was a measurement of that drift — not the models drifting, but the line between provable and unprovable shifting under her feet.
She opened the log. Five entries. Five overrides. One ATTESTOR-7 column showing the gap between human and machine judgment narrowing with each case, not because the machine was getting smarter but because the humans were getting less patient with the kind of doubt that could not be automated.
She added a sixth column. Left it blank. The header read: "What I cannot prove but cannot stop seeing."