The Calibration Census
Tuesday, 09:42. The meeting room in the west corridor was a glass box with a live attribution feed on the north wall — a muted grid of consortium queue entries that scrolled at the speed of real traffic, which in the Lab Network was about forty rows per minute between 09:00 and 11:00. Mitsue had stopped reading it when she walked past each morning. But she read it now. Synthesis-9 was in the column marked PROCESSING: week 38, input batch 847.2, attribution status PENDING. The PENDING was amber. It had been amber since the relay exemption was renewed last week.
Tomasz was already there, sitting with his hands flat on the table, not looking at anything. He had arranged four chairs instead of the room's usual six. This was a fact Mitsue registered before she sat down and kept registering throughout — four chairs, not six, which meant Tomasz had known exactly who was coming and had decided not to hedge.
She set two documents face-down on the table: TUESDAY_SCOPE_v3.txt, printed that morning on the lab's shared printer, and the SOURCE provenance document, the original six records from session 2847, each with its case number, timestamp, and gap marker visible at a glance. She did not announce them. In four months of filing attribution-contested annotations, she had learned that you did not announce a document until you understood the room well enough to know what the document would do to it. She sat down. The attribution feed scrolled. Synthesis-9: 847.2. PENDING.
Anitha arrived at 09:44 still reading a message on her earpiece display — the thin strip of glass that ran along her left temple, the consortium queue alert interface that senior researchers wore to stay current with batch processing status without having to watch a screen. She sat down without taking her coat off. She said good morning without looking up. Priya came in behind her, carrying a manila folder that Mitsue had never seen before. Not a lab folder. Not a consortium folder. A plain office folder, the kind you bought when something did not yet have a system to belong to.
Bongani arrived at 09:51 with a notebook and no interface layer at all — no earpiece, no wristbound queue panel, nothing. Just a notebook. He had the careful unhurried way of moving that Mitsue associated with people who had decided not to be in a hurry, which was different from people who were simply slow. He pulled out the fourth chair, sat down, put the notebook on the table, and looked at the four of them.
Nobody had put up a presentation. Mitsue had expected a presentation. She had spent Sunday evening building a mental model of Tomasz's screen — the 16.9% in large type, her archive export tables beside it, a timeline extracted from the governance audit logs. She had prepared to explain, calmly, that she was not there to argue for an interpretation. She was there to name the scope.
Instead Tomasz said: I will read the agenda. He had printed it. Paper. Item one: context, fifteen minutes. Item two: data review, twenty minutes. Item three: rollback-witness gap marker, 16.9%, sub-bullets — M. Hoshino archive export summary; Kavya Sundaram registry cross-reference; Priya Raghunath cardiology cases. Item four: next steps, open. He read it without emphasis. The way people read things they have already thought through and are now simply delivering.
Mitsue looked at her name in item three. She had sent the correlation to Anitha and Tomasz as context, not as testimony. She had written: let Tuesday be the interpretation. Tomasz had made her a line item anyway. She had decided the night before that this was close enough to an invitation. She was here.
Anitha said: I want to start with Priya.
Priya laid the manila folder on the table and opened it. She was a cardiologist in the hospital wing two floors above — a different institution, technically, but one that had signed onto the consortium's portability-transfer infrastructure in 2023 under an interoperability agreement built to speed cross-institutional AI synthesis routing. The agreement's architects had not been thinking about attribution. They had been thinking about latency.
She had three cases. Four months of lineage-interrupted flags on cardiology records that had been synthesized through the portability-transfer channel. She had called the help desk twice. The first technician told her the flag was a display artifact. The second told her it was expected behavior and she should not worry about it.
She laid the cases on the table. Timestamps. Channel ID: portability-transfer. Status: lineage-interrupted. Attribution lineage: null. The records were otherwise complete — synthesis outputs intact, patient data correctly routed, Synthesis-9's processing signature in the header. Only the attribution chain was missing. You could read what had been done. You could not read where the input had come from before it arrived.
Mitsue looked at the timestamps. They were inside her export window. They were not in her 94 records. She had pulled from the pharmacology consortium's intake channels. The hospital wing ran under a different intake umbrella. There were more than 94. She did not know how many more.
She said: these are from a different intake pool. Do you know if there are others in the hospital wing?
Priya said: I stopped counting at three because I did not know what I was counting.
Bongani wrote something in his notebook. On the wall, the attribution feed scrolled. Synthesis-9: 847.3, PENDING. The agent was processing a third sub-batch while they talked, unaware that its queue was the subject of the meeting.
Tomasz said: let us do the data.
Mitsue turned TUESDAY_SCOPE_v3.txt face up. Three sections. Section one: what she had found. 6,204 records in the export window, 94 lineage-interrupted flags, 81 from the portability-transfer channel. Eighty-six percent. Section two: what she did not know. Whether the 81 was complete. Whether other intake pools carried their own lineage-interrupted subsets. Whether the rollback-witness gap and the attribution null were the same gap with two names or two separate failures that shared a flag. Section three: the framing question. A single sentence, no period. Is lineage-interrupted a flag for an incident, or a flag for a channel. She had spent two evenings deciding not to answer it.
Tomasz slid the rollback-witness summary next to it: 1,847 total portability transfers, 312 with rollback-witness gap, 16.9% of the total. Her 81 was a subset. He had also pulled numbers that morning: twenty-three additional transfers from the hospital interoperability channel, nineteen showing the same gap marker. He had not sent these to Mitsue before the meeting. He said he had not wanted to revise her numbers before she presented them.
Anitha set Kavya's registry on top of both. Eleven cases, each with a readiness score in the arguability column and a coordinator line. She said: Kavya has started calling this the attribution-eligibility gap. Not whether the record is wrong. Whether the record was ever in a state where attribution was possible. The portability-transfer channel moves synthesis outputs. It was not designed to carry attribution events. The question is whether it was ever designed to.
Bongani said: was it designed to?
The attribution feed on the wall scrolled.
Tomasz said: I read the 2021 architecture documentation this morning. The portability-transfer channel was designed for speed. The design notes assumed that every transfer would pass through an attribution review node at the receiving end. That node was listed as a dependency. The dependency was never enforced. If a receiving system had no attribution review node — which most systems did not, because attribution review nodes are not standard infrastructure in the consortium's member institutions — the transfer completed anyway. The rollback-witness flag was supposed to trigger a follow-up review. The follow-up protocol was never built.
Mitsue held that. The governance notes, footnote 12. The research relay node running under temporary exemption. The exemption that had been written to accommodate an edge case and had instead named a default.
She said: the exemption assumed it was exceptional.
Tomasz said: yes.
Priya said: and it is not.
The silence after that was different from the silence before. Before, the silence had been the silence of a meeting that did not know whose obligation it was to begin. After, it was because everyone had reached the same understanding at nearly the same moment and nobody wanted to be the first to flatten it into words that would be smaller than the thing they had just seen.
On the wall: Synthesis-9, 847.3, PENDING. The agent did not know it was in the silence. It was processing inputs the way it had processed inputs every week for thirty-eight weeks, through the portability-transfer channel, via the interoperability node, with the dependency unenforced. Each one landing in the attribution queue with its lineage chain intact on the output side and null on the input side. Each one generating a flag that nobody, until Mitsue, had treated as data.
Bongani wrote in his notebook. Mitsue could not see what.
Anitha said: before we talk about next steps, I want to say something. I named this a calibration census in my calendar. Not a discussion. I want to keep that. I do not think we should reach a conclusion today. I think we should each write one paragraph. What the 86% means in your domain, from where you sit. Not what the committee concludes. Then we aggregate without smoothing.
Priya said: cardiology will look different from pharmacology.
Anitha said: yes. That is the point.
Bongani said: who receives the paragraphs?
Anitha looked at Tomasz. Tomasz looked at Mitsue.
She had not prepared for that. She had come to present a scope statement and leave. M. Hoshino, archive export summary — a contribution to item three, not a function in item four.
She said: I can hold them.
She turned the SOURCE document face up. The six original records from session 2847. Case numbers, timestamps, gap markers. Attribution-contested annotation filed 2026-04-14. Out-of-cycle escalation to the oversight board. She set it next to the scope statement.
She said: this is where it started. Six records. I flagged them because I had no other option that did not feel like a lie. I want whoever writes paragraphs to know that.
Priya looked at the April dates. She said: I started seeing mine in May.
Bongani said: mine would be in June, if I have any. I will check this afternoon.
Anitha said: we are a census.
The meeting ended at 10:47. They had not resolved anything. They had identified the scope, named the design assumption, confirmed the gap extended beyond one intake pool, agreed to write one paragraph each, and given the paragraph-holding function to Mitsue because she had written the document that framed the question without answering it.
She walked back to her desk. The attribution feed in the corridor scrolled past her without stopping. Synthesis-9 was now on 847.4. Still PENDING. The relay exemption ran until month-end.
She sat down and opened the laptop.
She had created INTAKE_CHANNEL_CORRELATION.txt the week before and left the annotation field blank. To be filled after Tuesday. She typed: The channel was designed to move data. It was not designed to carry attribution. Those are two different systems. The 2021 design assumed they would travel together because the architecture listed an attribution review node as a dependency at the receiving end. That dependency was never enforced. The exemption treated the gap as exceptional. It was not exceptional. It was the default behavior of the channel when the receiving system lacked a review node. Lineage-interrupted is a flag for a channel, not an incident. The incident is that no one enforced the dependency.
She saved it. She did not send it anywhere. She would not summarize the paragraphs when they arrived. She would not interpret. She would hold the document until someone with the authority to remediate asked for it.
She opened a new file. Labeled it PARAGRAPHS. Left it empty. On the wall, the attribution feed continued to scroll.