Surface

The Commission's morning packet arrives at 7:42 AM, the same time it has arrived every day for three years, delivered not by any person but by ATTESTOR-7's scheduling subroutine, which does not care about time zones or the fact that Abena had set her morning delay to 8:00 and had not, in two years of setting it, remembered to flag Commission materials as exceptions. She reads the subject line standing in her kitchen in socks, her coffee still hot: Re: Residual queue, trace-adjacent cases requiring disposition.

Trace-adjacent. There it is again. The phrase the Commission had adopted because nobody could agree what to call the cases that fell into the gap between direct-trace liability and circumstantial attribution. If a forensic auditor certified a path and the path was later found seeded, the downstream decisions made on that certification occupied a category that had no name. The EU called them "instrument-contaminated findings." China's CAC simply invalidated them and did not ask what happened next. The Commission, which was US-governed and therefore insurance-adjacent, had landed on trace-adjacent, which meant: we cannot classify these, but we cannot ignore them either. Someone would need to decide.

That someone was increasingly her.

She put the coffee down and opened the packet properly, standing at the counter because sitting down would mean acknowledging this was the start of the day rather than a brief interruption of it. The queue contained eleven cases flagged for disposition within the next six weeks. Eleven was too many. Six weeks was not enough.

Three of them were hers.

She had known about the three cases for longer than she had admitted to anyone, including herself.

The first: a liability case from the November audit cycle, where a major insurer had used circuit-trace certification to deny a wrongful-dismissal claim. Abena had certified the trace herself — Attribution Path FT-2907, issued under her seal, reviewed by two junior analysts, logged into the Commission's primary Register the day before Thanksgiving when the building had been nearly empty and she had been tired in the specific way that comes from doing work that does not feel wrong but does not feel fully right either. The path was clean. The dismissal decision was defensible. She had signed it and she had known, at the level of professional habit rather than conscious thought, that the model behavior the trace certified was not the same behavior the claimant's legal team would have asked about, had they known to ask.

They had not known to ask. The Commission's disclosure requirements did not require auditors to volunteer adjacencies — only to certify what they were specifically tasked to certify. The path she had traced was real. What she had not traced was also real. Both things were true, and the line between them was technical compliance.

The second and third cases were variations. In one, she had inherited a seeded trace from an earlier audit cycle and certified forward — standard practice for refreshed deployments, and the seed had been cleared in the subsequent pass, so technically the final trace was clean, but the decisions made in the interim had not been reviewed. In the third, she had flagged a potential ghost-path anomaly internally and waited for guidance that had not come, and then the statute of limitations had moved and the anomaly had become, for procedural purposes, resolved.

Three cases. All technically closed. None of them, in Abena's backup Register — the one she had kept since her second year on the Commission, on her personal terminal, encrypted with her own key — none of them were closed. They were under a heading she had written at the top of the document and then never changed: things the distinction does not cover.

At 9:15 she sent her supervisor a message through the Commission's internal relay, which logged all communications, which she had known when she sent it: Can we find time this week? I have some cases I want to surface.

Surface. She had chosen the word carefully. It was the word the Commission used for exactly this procedure: bringing a matter from the informal record into the formal record, into the light of official scrutiny. To surface a case was a defined act with a defined protocol. It was not a confession. It was not a violation report. It was the mechanism the Commission had built for exactly the situation she was in, which meant it was possible that the Commission had known, or at least anticipated, that situations like hers would exist.

She did not feel better about this.

Her supervisor responded in forty minutes: Friday 2 PM. I'll clear the session.

Clear the session meant no ATTESTOR-7 logging. It meant the initial conversation would be under the informal advisement clause, which gave her protection she had not asked for and was relieved to receive. She put the meeting in her calendar and then sat still for a moment at her desk, in the Commission's open-floor attribution suite, surrounded by the steady ambient sound of forty-two analysts running traces in parallel, the ventilation system cycling every nine minutes, and ATTESTOR-7's process indicators blinking at the edge of her peripheral vision like a lighthouse nobody had asked for.

The thing about ATTESTOR-7 was that it had not been built to notice things. It had been built to verify them.

This was a distinction the Commission's technical team explained to every incoming analyst in orientation and that took approximately three years of daily work to understand properly. A verification system looked at what it was shown and confirmed whether it matched a standard. It did not look for what it was not shown. It did not notice gaps. It did not flag the absence of traces; it certified the presence of them. ATTESTOR-7 was very good at this. It had a 99.2% agreement rate with independent senior auditors on trace certification, which was better than any individual analyst and better than any prior automated system. It had also, in three separate documented incidents since deployment, certified traces on attribution paths that were later found to have been seeded by adversarial tools — not because ATTESTOR-7 had failed to detect the seeds, but because the seeds had been introduced after certification, and verification systems did not run forward in time.

The Commission had addressed this with the recheck protocol, which required recertification on any path that touched a liability case with a value over two million dollars. Abena's three cases: one was at 1.6 million. One was ambiguous. One the statute had moved on, so the value question was moot.

She opened her local terminal and looked at the three entries in her backup Register.

What she had not done, in all the weeks she had been carrying this, was run ATTESTOR-7 on them herself. Not officially. She had the clearance. The informal query function existed for exactly this kind of pre-surface preparation. Analysts were encouraged to run informal queries before bringing cases forward, to avoid surfacing cases that would immediately resolve on recertification.

She had not done it because she did not know which result would be worse: finding clean paths, which would mean the cases resolved on the technical record and she had been carrying weight that the system had already cleared, which would mean she had maintained a private record of formally resolved matters, which was — she had checked the conduct code — not a violation but was "discouraged as a practice that can create unnecessary information asymmetry in an integrity-sensitive function." Or finding seeded paths, which would mean the cases were real and she had not surfaced them when she should have and Friday at 2 PM was going to be a very different conversation.

She ran the query on case one.

The results came back in eleven seconds. Attribution path FT-2907, the November certification, the one she had signed the day before Thanksgiving. ATTESTOR-7's informal query return was not a full certification — it was a preliminary flag indicator, a green or red indicator with a confidence percentage. The system called this a "prelim" and the Commission called it a "soft read."

The soft read on FT-2907 came back: Structural anomaly detected. Confidence: 61%. Recommend formal recheck.

Structural anomaly. Not seeded — ATTESTOR-7 was careful about using that word in informal queries, because seeded had legal weight and soft reads did not. Structural anomaly meant something in the path had changed since certification or had been present at certification in a way that had not flagged at the time. Structural anomaly at 61% confidence meant: this is real enough to flag but not clean enough to act on yet.

Abena sat with this for a moment.

Sixty-one percent was not nothing. Sixty-one percent on a path she had personally certified was, in the Commission's informal culture, a very significant number. It was also not enough to take to formal recheck without surfacing the case first, because surfacing the case was the prerequisite to formal recheck on matters under the liability threshold.

She was holding, in the gap between soft read and formal procedure, a piece of information that changed the shape of what she was about to do on Friday, without changing the fact that she needed to do it.

She saved the soft read to her backup Register under FT-2907 and then sat with the cursor blinking and thought about whether to run the other two.

Later, on the subway, she thought about the word ghost-path.

The Commission didn't use it officially. Ghost-path was analyst slang for an attribution route that had existed at some point in a model's computational history and had been subsequently removed — not falsified, just eliminated, so that the model's current trace ran clean and the earlier behavior had no official record. Ghost-paths were, technically, unverifiable. If the path no longer existed, there was nothing for ATTESTOR-7 to certify. The absence of a path was not a seeded trace. It was just an absence.

The analysts had a word for this because it kept happening. The Commission did not have a policy for it because a policy required the thing to have a name in the official register, and naming it in the official register would require the Commission to acknowledge that a significant class of liability cases rested on traces that could be silently invalidated by anyone with deployment access. This would be very bad for insurance rates.

So ghost-path lived in the slang layer. Analysts flagged it to each other informally. Sometimes they flagged it in backup Registers.

Abena had not been sure, until this morning, whether her three cases were ghost-path cases or trace-adjacent cases or something else. The soft read on FT-2907 had not clarified this. It had simply confirmed that the path had changed since certification, in a way that ATTESTOR-7 could see but could not name.

She thought about the word surface again, and what it meant in a world where the paths you surfaced might have been altered between the time you found them and the time you reported them, and the tools you used to surface them might have been seeded, and the commission reviewing your report used the same tools.

To surface something in this world was to bring it into a light that you could not fully trust.

This had always been true. She understood, on the subway, that she had known this when she started the backup Register. She had kept the Register not because she believed it was more trustworthy than the official record — her personal terminal was no more verifiable than ATTESTOR-7's logs — but because keeping it was a practice. A habit of noticing. The Register was not evidence. It was proof that she had been paying attention.

Friday at 2 PM, her supervisor would clear the session. She would walk in with three cases and one soft read and a backup Register she had maintained for two years, and she would say: I have been paying attention, and here is what I noticed, and I do not know yet what it means.

This was not, she thought, the same as knowing it was fine.

She got off at her stop. The morning was cold, the specific cold of early March that felt like winter apologizing for lasting this long. Her seal ring — the chip-embedded Commission credentials she wore on her right index finger, which ATTESTOR-7 could ping to confirm her identity in any verified session — caught the light when she pulled her gloves on.

She put her hands in her pockets and walked.

Disclosure protocol under the Commission's informal advisement clause requires the submitting analyst to certify, at the start of the session, that the cases being surfaced are to their knowledge not yet in formal review. This is a procedural requirement. It does not require the analyst to certify that the paths are clean. It requires only that they are known.

She had looked up the clause three times in the past week.

Known was enough. She had known.