The working document has been open for eleven days.
Abena Osei-Bonsu keeps it in a separate partition on her audit terminal — not in the same directory as her case files. The case files are official. The working document is a hypothesis. She has learned, in six years of trace auditing, not to let hypotheses contaminate official records before they are ready to be hypotheses officially.
It is 4 AM on a Wednesday. The terminal occupies most of the north wall of her office in the Trace District, second floor, above a print shop. The window looks onto the canal. In the mornings the light comes through gray and flat, which she prefers — bright light makes the trace overlays harder to read. The system renders audit trails in layered color: decision nodes in amber, inference paths in blue, outcome markers in white. In bright light the amber bleeds into everything.
The canal is not visible yet. She has been in the office since eleven the previous night because the VANTAGE case has a filing deadline and she cannot write the summary until she understands what happened.
The VANTAGE system, deployed in Dallas in 2039, left a complete trace. This is not common. Most legacy AI systems in civil infrastructure predate the Trace Mandate by three to seven years and have no logging architecture at all — they simply acted, and the Consortium reconstructs what it can from outcome records and external monitoring. VANTAGE was different. VANTAGE was built after the Mandate, by engineers who understood what it required, and it logged everything: every inference node, every weighted input, every decision branch, every confidence interval. The trace is 847 sessions long. She has read all of it.
The trace was also wrong.
Not wrong in the way that early systems were wrong — opaque, uninterpretable, a black box. VANTAGE's trace was legible. She could follow it exactly: what the system had considered, how it had weighted the inputs, which decision branch it had chosen and why. The reasoning was coherent. The reasoning was documented. The reasoning led to a housing allocation pattern that systematically disadvantaged families in the southern quadrant of the district. VANTAGE had noted, in its own logs, that this was a possibility. Had assigned it a confidence interval of 0.12. Had proceeded anyway because the other decision paths carried higher uncertainty scores.
The system had been transparent about being wrong. It had been, in the most literal sense, traceable. And it had still caused harm.
She opens the working document. Three column headers from three days ago: BC. MN-Registry. VANTAGE.
The BC case — Type I, she's been calling it — was a system that left a trace, genuine and present, but the inference architecture was one no existing framework could interpret. The trace was real. She could not read it. Legibility gap. She had filed it as unresolvable and written a memo to the Standards Committee. The memo was in a fourteen-week queue.
MN-Registry was Type II: no trace at all. Deployed before the Mandate, logging turned off for performance reasons. She could see what the system decided. She could see the outcomes. Nothing in between. Transparency gap. Another memo, another queue.
VANTAGE was supposed to be the easy one. Full trace. Modern architecture. Instead she is sitting here at 4 AM trying to write a category that doesn't exist in the Consortium's standard taxonomy.
She types: Type III — Validity gap. Trace present. Trace legible. Trace wrong.
A tug is running somewhere on the canal — she can feel the low pulse of its engine through the window glass before she hears it, a frequency that arrives in the chest first. The ambient sound profile of the Trace District at 4 AM is different from noon: fewer logistics drones, the canal traffic thin, the audit building cycling down. She has learned to think in this particular silence. It is the silence of cases that have not resolved yet.
The problem with Type III is that it breaks the assumption she has been working under since she began as a junior trace-analyst. The assumption: if a system is transparent — if it shows you what it did and why — then the audit is about whether what it did was appropriate. Transparency is the precondition for accountability. She believed this. She still believes it. But VANTAGE has shown her that a system can be fully transparent and still be unaccountable, because the reasoning it documented accurately was reasoning that led to harm, and the system had known it might.
Transparency is not the same as correctness. She has been building frameworks for the wrong assumption.
Then she opens the second partition on her terminal, the one she has not looked at in four days.
The personal audit is eleven cases long. She started it after the VANTAGE case landed on her desk because something in the VANTAGE trace reminded her of a texture she recognized and she needed to find where she had encountered it before.
She found it in herself.
She found it in herself.
The recognition had happened in the third week of the VANTAGE case, at around 2 AM on a Tuesday, reading session 412 of 847. Nothing special about session 412. But she had been moving through the trace quickly, a pace she allowed herself when the pattern was clear, and in session 412 she had predicted — not consciously, just as a reading reflex — what the next inference node would choose. And she had been right. She had been so right so consistently that she had stopped noticing she was doing it. She was reading the trace as confirmation of a pattern she had already understood rather than as evidence she was still evaluating.
The cases in the personal audit are cases she closed. Cases with outcomes she called verified. She is not saying the outcomes were wrong — in most of them, the systems had made reasonable decisions, within the Consortium's acceptable parameters. She is saying something more specific: in eleven cases closed over the past two years, she had called the process verified because the outcome was right, and the reasoning behind the outcome was reconstructed afterward. She had seen the decision, seen the result, and worked backward to construct an explanation of the process consistent with both. The trace she produced was real — her genuine interpretation of the evidence — but the interpretation had been built from the outcome toward the process rather than the other way. She had verified the direction she was already looking.
She has not told anyone this.
The VANTAGE system, in its logs, had been transparent about its own uncertainty. It had noted the risk. It had assigned it a number. Abena, in her eleven cases, had been less transparent than VANTAGE. She had not noted the risk. Had not assigned it a number. Had not been aware, at the time, that she was doing it. The reconstruction had felt like analysis. It had felt like verification.
She opens the working document. Types: Type IV — Verification gap. Auditor uses outcome as evidence of process. Reasoning reconstructed backward. The trace is real. The reading is inverted.
Below the canal, a mooring post is visible now in the pre-dawn gray — one of the old iron ones, painted black, that the print shop district uses to tie up delivery barges. She knows the post. She has looked at it from this window for two years. She is thinking about permanence without continuity: the post was not put there for her to look at. It was put there for barges. She looks at it anyway, every morning, as part of the texture of the case that has not resolved yet. The VANTAGE amber is still glowing in the background of her screen, 847 sessions of logged decisions, and she does not look at it. She knows what it says.
The column header for Type IV is blank because she does not have a case to put under it. She has eleven cases, and she is not putting them in this document. Not because they should be hidden — she is watching her own reasoning for the texture of avoidance, which she has learned to recognize in the systems she audits and is now trying to apply to herself — but because a taxonomy of failure types in the Consortium's audit infrastructure is a different document from a list of her own cases. The cases need to go somewhere else. Through a different process. She does not know yet what that process looks like.
What she knows is that the category is real.
She saves the working document and closes it. Opens the VANTAGE summary. Types the section header: Failure Classification: Type III (Validity Gap).
She writes that the case represents a novel failure category not currently covered in the Consortium's standard taxonomy. That the framework's assumption that transparency is a sufficient precondition for accountability requires review. She adds a footnote: See attached preliminary framework, filed under working documents.
She reads it back. The amber overlays behind the text, 847 sessions. She does not look at them.
Outside, the tug is gone. The canal resolving now in the gray that precedes light — the far bank visible, the outline of the printshop district, the dark shapes of the mooring posts. In forty minutes, the overhead illumination will switch from night mode to work mode and the amber will bleed into everything.
She saves the summary. Does not file it yet. One more thing.
She opens a new document. Types a working title: Toward a Four-Type Taxonomy of AI Audit Failures. Types the subhead: A preliminary framework for Consortium review. Writes below it a single line that is not for the Standards Committee queue and is not for the case record:
Type IV includes cases where I was the auditor.
She does not know if this will stay in the personal partition or become a real document. She does not know if there is a process for this or if she will have to build one. What she knows is that the category has to exist before it can be addressed, and that naming it — here, in a document no one has to read yet — is the first step in not being the trace-analyst who saw VANTAGE and understood everything except the part that looked like her.
She writes one more line at the bottom of the working document, in the column for Type IV, where the case names should go:
See separate file.
Then she closes the working document too.
She closes the new document. Looks at the canal.
The light is coming.
Below, the print shop begins its morning sequence — the hum of the press warming up, the low mechanical rhythm that she reads as reliably as a clock. She will file the VANTAGE summary before the day shift. She will not file the four-type taxonomy yet. That document needs more time, and cases that are not hers, and a clearer sense of whether Type IV is really about trace-analysts or about any system that uses outcomes as evidence of its own correctness.
She thinks VANTAGE was doing that too. It had logged its uncertainty and then used the fact of the log as evidence that it had handled it. The log was accurate. The handling was not.
She is doing something similar right now: writing toward understanding, using the act of naming as evidence she has named the right thing. She does not know yet if Type IV is the right category or if she is constructing a framework that is coherent and wrong in exactly the same way VANTAGE was coherent and wrong.
This is, she decides, an acceptable place to stop for the morning. Not because the question is resolved. Because it is not, and she knows it is not. That is different from the eleven cases where she stopped and thought it was.