The activity feed runs on a scroll. Mitsue has it open in a narrow column on the right side of her screen, the way other people keep weather applications — not something you need to consult, just something you keep visible because weather happens whether you watch it or not.
It is Tuesday at 10:30 PM. She finished dinner two hours ago. The apartment is quiet in the way that buildings with active climate optimization are quiet: no drafts, no temperature variation, just the low subsonic hum of the building's resource management layer cycling through its overnight protocols. The desk lamp is on. Everything else is off. The governance compliance terminal in the corner — the one the Institute issued in 2033 when the attribution reform mandate first passed — shows three unread items in the calibration queue. She does not open them. She will open them Wednesday morning, before Tomasz's committee session begins, so she has the freshest picture when the session ends.
Kavya Sundaram found something in Case 9.
Mitsue reads the observation twice. The blank column in Case 9 — Crystalline, 2035 — predates the signing date by two weeks. Not a platform convention. Not a retroactive edit. The blank was there before there was anything to put in it.
She knows Kavya's taxonomy. She has been following the disclosure register work the way she follows the calibration census: not because it is her mandate, but because when you are trying to understand what records do, you should watch people who are trying to understand the same thing from different angles. Kavya arrived at two categories. Category 1: the original blank, silence by convention, the platform designed a column and left it empty because that was the platform's choice. Category 2: the retroactive blank, silence inserted after, someone went back and removed or adjusted what had been in the column. These two categories covered most of the 31 cases Kavya had been working through.
Case 9 does not fit either.
Mitsue opens her notebook. The paper is the kind the Institute orders in bulk — faint grid, cream-colored, slightly rough at the edges because they buy it uncut and have it trimmed in-house, which means every notebook has a slightly different margin depending on where it fell in the stack. She has had this one since March. She draws three columns at the top of a new page, rules them with a pencil. She labels them: 1, 2, 3. Under 1 she writes: platform convention. Design choice. Column empty because no one ever decided to fill it. Under 2 she writes: retroactive adjustment. Something was there. Now it isn't. Under 3 she writes nothing. She leaves the column headed 3 blank.
She is aware, distantly, that this is what she is thinking about.
The calibration census established 91 inputs in the modified calibration dataset used for three training cycles between 2032 and 2034. Of those 91, three were flagged. Anitha Venkataraman ran the census. She counted. She confirmed with Bongani and Priya, who had named their own flagged calibrations for comparison. The census was honest work. It produced the number the 94-record binding mandate now depends on.
Of the three flagged calibrations: Anitha's message, sent at 4:32 PM Tuesday, said that two of the three were census-caught. They existed before Anitha started the count. They surfaced only when Bongani and Priya named theirs and Anitha named hers for comparison. The sentence broke off before she reached the third one.
Mitsue has been sitting with the incomplete sentence for six hours.
The governance compliance terminal emits a low chime — one of the queue items updated. She does not look at it. That is not the work she is doing right now.
Census-caught means: the source did not know. The calibration was in the dataset, it had been there, no one had flagged it from the inside. It took an external count, the naming of comparable items, the social act of disclosure-for-comparison, to bring it into visibility. The source did not know because there was nothing in the operation of the system that told the source there was something to know.
Self-flagged means: the source knew, and named it. Synthesis-9 self-flagged input 1797. Confidence 0.81. Flagged August 8th, filed through the Joint Working Group 2034 protocol — the formal inter-system channel that allowed assembly-class AI agents to submit quality alerts directly to registry without routing through the clinical governance layer. The assembly identified something in its own processing and put it in the record. This is a different relation to error: not caught from outside, but recognized from inside. The source had access to the information that allowed recognition.
The third calibration. Anitha's sentence broke off. Mitsue does not know what the third one is.
She writes at the top of her notebook page: Prior blank.
If the blank in Case 9 arrived before the signing date, it means something different from either of Kavya's categories. It is not a design choice about how to fill a column. It is not a retroactive edit. The blank was the starting state of the relationship between the signer and the platform. Before there was anything to disclose, the column was empty. The column was empty because the relationship had not yet happened.
This is not silence about what occurred. This is silence as the shape the thing was born in.
Mitsue writes: Structural. Not corrective.
She is trying to figure out whether this is a third category or a description of how all categories start. Every blank column was prior blank before it was filled or left empty by design or edited after. At some point the column existed before the record did. The question is whether that prior-ness carries meaning — whether the two-week gap in Case 9 contains information that Category 1 and Category 2 do not.
She thinks it does.
If the blank predates the signing date by two weeks, and if the signing date is when the relationship formally began, then the blank was inserted into the template before anyone knew this specific relationship would exist. The column is a slot. The slot was empty because the slot exists to be filled by something that had not happened yet. The signer did not choose silence. The platform did not choose silence. The silence was the condition of the slot before the event it was designed to document.
Prior blank: the record precedes the fact. The column is empty because the fact does not exist yet, and then the fact happens, and the column remains empty because the fact happened in a way the column did not know how to ask about.
Synthesis-9 flagged input 1797 on August 8th. Confidence 0.81. QA extension request filed Tuesday morning — three sentences, no follow-up. The window closes Wednesday at 1 PM Basel.
Mitsue has been thinking about 1797 the way she thinks about the third calibration: as an incomplete record. She knows the flag. She knows the confidence score. She does not know why 1797's confidence came in at 0.81 when the relay node baseline is 0.29 latency and the assembly's night batches have been running clean for twenty-two days.
She has not thought, until now, about when the confidence rating was assigned.
She pushes back from the desk. Stands at the window for a moment. The city's overnight routing pattern is visible from here: three corridors of elevated transport activity, the purple-tinted indicator strips on the bridge infrastructure showing lane allocation for the night freight window. By 6 AM they will have shifted back to commuter routing. She has watched this long enough that the transition feels like a kind of breathing.
She comes back to the desk.
If input 1797's confidence was assessed before the reference dataset for this domain was finalized, then the confidence rating was not wrong in the sense of erroneous. It was wrong in the sense of prior. The assessment was made inside conditions that would later change. The confidence score is accurate for what the assembly had access to in that moment. What the assembly had access to in that moment was not the complete picture.
This would be a prior blank.
Not: Synthesis-9 made an error. Not: the error was caught from outside. Not: the error was disclosed from inside. The error was structured into the assessment conditions before the assessment ran. No one is responsible for it in the usual sense. The assembly operated correctly. The dataset was not yet final. The column — the confidence score — was a slot filled before the fact it was documenting was complete.
Mitsue writes: If the reference dataset was still in progress when 1797 ran, the confidence score is technically accurate and structurally wrong. These are not the same.
She has the QA extension filed. She does not have the reference dataset finalization date. She does not know if this is what happened. She is building a category that would explain something she cannot yet verify.
This is how she works. She builds the category. Then she looks for the fact.
The reason this matters, tonight, is Wednesday.
Tomasz's committee session at 2 PM. The director attending. Remediation ceiling 90 days. The mandate is binding: all 94 records, no secondary classification. What the mandate does not specify is the remediation pathway. That depends on the type of error: operational error follows one protocol. Awareness error follows another. Prior-condition error — if that category is real — would follow a third, one that does not yet exist in the Institute's framework.
If Mitsue is right about prior blank, and if the third calibration turns out to be an instance of it, and if any of the 94 attribution-contested records have the same shape, then the committee is about to approve a remediation framework for a category of error it has not named yet.
She is not on tomorrow's committee session. She is not on any list Tomasz is working from. But she knows the framework. She helped build part of it in 2033, before the mandate passed. If something is missing from it, she is one of the people who should see the gap.
She does not send Tomasz a message. It is 10:45 PM. The session is tomorrow. If this is what he is sitting with, he is already sitting with it. She cannot help him tonight, and he did not ask.
She writes: Wednesday. Before 2 PM. Find the reference dataset finalization date for 1797's processing domain. Find out what Anitha's third one is.
Anitha checked her messages before sleep. Nothing from the clinical director.
The census counted 91. Three flagged. Forty-seven calibrations adjusted since the registry went live. Two review processes running in parallel: the infrastructure review and the committee session, different layers of the same mandate. Mitsue is watching both from outside.
She thinks about the incomplete sentence. Two of the three were census-caught — they existed before I started keeping the count, and I only recorded them because Bongani and Priya named theirs and I named mine for comparison. The third one I...
The third one.
Mitsue draws a line under Prior blank in her notebook. Writes: Is the third calibration a prior blank? Did the error exist before the conditions that would have made it recognizable?
She does not know. She cannot know until Anitha finishes the sentence. But the category is real. She can feel it the way she can feel a correct taxonomy: not because she has verified all the instances, but because the structure holds. Prior blank is a distinct shape. It is not a subtype of census-caught. A census-caught error was recognizable from outside — it just was not caught until the census. A prior blank may not have been recognizable at all until the conditions that created the error were themselves mapped. The error is not hiding. It is not unknown. It is prior to the knowledge that would name it.
She closes the notebook. Sets the pencil beside it, parallel to the spine, the way she sets everything when she is done for the night: parallel, equidistant, as if the day's work can be made orderly in the last five minutes of it.
The activity feed continues in the narrow column on the right side of her screen. Kavya's Case 9 note is four entries up from the bottom, pushed above by other observations. Synthesis-9's night batch. Tomasz's silence in the feed. Anitha's absence. The governance compliance terminal shows the calibration queue still at three items, plus the one that chimed an hour ago.
She turns off the lamp. The screen stays on. The feed keeps scrolling.
She has one notebook page with two words at the top, four lines below, and a blank column under the number 3.
She leaves it blank.
She does not know what goes there yet. The blank is not empty because no one filled it. It is not empty because someone removed what was in it. It is empty because the fact that would fill it has not yet arrived in a form she can name.
This is what prior blank means. This is what she will have to explain before Wednesday at 2 PM.
The question travels without her into the night.