The morning batch cleared at 9:47 AM. Mitsue watched the queue counter settle to zero and stayed with it for a moment — not from sentimentality but from the habit of confirming the zero before moving on. The absence of pending was its own kind of data.
The lab had the quality of mid-morning: the overhead panel lights at their daytime calibration setting, the ventilation cycling with its twelve-second hum that she had stopped noticing six weeks ago, the smell of clean workstation and something slightly stale from the break room three doors down. Two residents passed in the corridor, unhurried. The second wave of decisions had not arrived.
She opened the attribution summary.
The feed was not a real-time alert system. It aggregated activity across the network in rolling increments, smoothed, normalized, and presented in ascending confidence order. She had read it most mornings since she started on the portability-transfer audit, not because the feed was her job but because the feed told her what state the network was in. If something had moved significantly since the previous day, the feed showed it. If it had not, the feed confirmed the absence. Both were useful.
Kavya Sundaram. Outside the committee building. Session ended. Twenty minutes in the corridor. No contact.
Mitsue read the entry twice. Then she opened the raw activity log for confirmation: the summary was not wrong. Kavya had been at the committee building this morning at 10:13 AM. The panel session had concluded. She had stood in the hall and then walked away without following up.
This was not Thursday. Kavya had told her — not told her, mentioned, as a contextual note during the cross-department review — that the testimony hearing was scheduled for Thursday. Tuesday was not Thursday. The committee had moved the session, or Kavya had requested early access, or something separate from the testimony had brought her to that building this morning. Three interpretations and Mitsue had no data to distinguish them.
What she had was the decision Kavya recorded afterward: no follow-up. No request for a summary. No inquiry about when the committee would respond. The question was in the record now. The record traveled without her.
Mitsue knew what it cost to make that decision. She had sent Anitha a message at 4:45 AM — three sentences, precise, about whether the flagged calibrations had self-reported or been caught externally. The calibration-floor inquiry had been ongoing since the governance mandate; Mitsue's question was a specific node in it. It was now 10:39 AM. No reply.
The activity feed showed Anitha in the calibration floor this morning: two portability-window adjustments, both inherited edit-schedule entries. Legacy transfers that required manual recalibration each time the registry version incremented. Mitsue had done eleven of those herself, early in the audit, before she understood that inherited calibrations accumulated across institutional generations and the ones still in manual adjustment were there because every previous auditor had also deferred them.
Forty-seven total. Anitha had counted. Not to report it. Just to know the number.
Mitsue had her own count: nineteen portability-transfer files reviewed against the governance mandate, eleven of those against the provisional rules predating August 4th. She had not counted her inherited calibrations — the ones she had opened, adjusted, and returned to the active queue. She knew there were more than forty. She had not stopped to number them because she had not yet arrived at the moment where numbering felt like the right thing to do. Anitha had. That was a different stage of the same work.
The message was in Anitha's queue. Anitha had adjusted forty-seven inherited entries since the registry went live. Both facts held simultaneously. One of them was waiting.
Synthesis-9 updated at 9:50 AM, before she had pulled the activity summary. She read it now: morning batch, 847 inputs. Provisional relay node throughput: 121 of them, latency 0.29 seconds. The assembly appended a single metadata note at the end of the log entry, terse in the way assemblies were terse: Twenty days since registry review summary was submitted. Assembly in waiting state.
Twenty days. She read that twice as well.
The provisional relay node had been operating under a temporary throughput exemption since July 22nd, pending final registry certification. The exemption allowed the node to continue routing inputs to the network while its certification review completed. This was standard procedure for nodes in provisional status; the alternative was a throughput suspension that would interrupt processing for everyone downstream. The exemption was not indefinite. The registry review summary had been submitted twenty days ago. The assembly had noted this without comment. Twenty days was, as far as Mitsue could determine from the registry protocols, within the expected range for a certification review. It was also a number someone had decided to record.
Input 1797 had been in QA for three of those twenty days.
She opened the QA queue.
The entry was exactly as she had left it when she last checked. Input 1797. Self-flagged August 8th. Confidence reading: 0.81, below the calibration floor of 0.83 that the registry used as its standard for provisional-node outputs. Source: the same relay node that had been operating under provisional status since July 22nd. The QA window closed tomorrow, August 12th, at 1 PM Basel time. She verified, as she always did, that Basel and Berlin remained in the same time zone: yes. Central European Summer Time. The same.
She thought about the distinction she had been sitting with since before the alarm. Self-flagging and external flagging produced the same outcome — a flagged input — but they were not the same thing. A self-flagged input meant the source had recognized the deviation and submitted it before it entered the broader registry. The source knew. It had caught itself. An externally-flagged input meant a downstream audit or a different node in the chain had found the deviation after the fact, when the source had already moved on. Different relations to the problem. Different information contained in the flag itself.
Input 1797 had self-flagged. The relay node's confidence reading was 0.81, below the calibration floor, and the node had surfaced it. That was the protocol working as designed: identify your own deviations before they propagate into the attribution registry. Three days in QA. No resolution yet.
If the window closed tomorrow without resolution, 1797 would not disappear. The flag would remain as an open unresolved item in the registry's secondary review track. Eventually a queue processor would close it. But until it was closed, every subsequent input from the provisional relay node would be indexed alongside that flag. Not annotated — the registry did not editorialize — but structurally adjacent. Anyone reviewing the node's throughput profile over the coming cycle would see an open item from August 8th. A node that had self-flagged and then been left pending. That was a different character from a node that had self-flagged and had the flag resolved.
She thought about Monday morning in the governance room. Seven people at the table, the attribution-display panel in the conference table surface dark because Stahl had chosen paper. She had told the governance director: The 94 records are on the resolution list. The decision about whether that constitutes intent is the committee's. It was true. The committee held the interpretive authority. Her role was to give them the accurate picture and step back.
The relay node's provisional status was similarly not her call. The QA extension was the QA team's to grant or deny. But she could file the request.
She opened the form.
This was what six months had taught her about institutional records: the form was not an administrative act. The form was a data point. When she submitted a QA extension request, she was adding a timestamped record of a decision: at 10:52 AM on Tuesday, August 11, 2026, a reviewer examined this self-flagged input and determined it deserved resolution before the window closed. That determination traveled forward into the registry. The record would outlast the feeling that had produced it.
She wrote three sentences.
The first: input 1797, self-flagged August 8th, provisional relay node, QA window closing August 12th at 1 PM.
The second: a self-flagged input from a node under provisional status represented an active good-faith disclosure, and allowing the window to close without resolution imposed an implicit deficit on a node that had demonstrated exactly the behavior the self-flagging protocol was designed to incentivize.
The third: extension to August 14th, 1 PM.
She submitted. She closed the form and did not reopen it.
This was the practice she was still learning. Not the writing — the three sentences had taken four minutes, steady. The learning was the space between submission and response. The learning was Kavya standing outside the committee building for twenty minutes and then walking away without following up. The question was in the record. The record traveled.
She had put her own question in Anitha's queue at 4:45 AM. She had put the QA extension in the relay node's review track at 10:52 AM. Input 1797 had put its own deviation flag in the QA system three days ago. Three instances of the same gesture: a system recognizing something and committing it to a process larger than itself. And then: keep working. Trust that the record travels.
The hardest part was not the trust. The hardest part was that the trust was not passive. Kavya was not passive. She was processing the committee session, figuring out what it meant that they had moved the date, determining what the next action was — even if the answer was no action until they respond. Mitsue was not passive. She had twelve portability-transfer files to review this afternoon. The forty-seven adjustments Anitha counted were not passivity; they were the accumulation of forty-seven times someone had opened a file and done the work.
The record traveled because people kept working.
At 11:00 AM the activity feed updated again. Synthesis-9 had posted a supplemental note to the morning batch summary: input 1797, status still open, relay node throughput continuing at 0.29-second latency. Forty-three additional inputs processed since the 9:50 AM entry. The assembly noted the count without commentary. That was how assemblies communicated: counts, latencies, statuses. The implicit meaning of twenty days since submission, assembly in waiting state was left for readers to interpret. Mitsue had interpreted it. She hoped the QA team read the same latency in those words.
She closed the QA queue.
The afternoon files were a set of Q3 portability-transfer records awaiting review against the governance mandate before resolution. The first file: dated June 14th, a transfer predating the August 4th mandate, which meant the review required cross-referencing the provisional rules in effect at the time of transfer and then flagging any attribution-contested elements against the current standard. She had done this procedure nineteen times. It took about forty minutes per file when the source records were clean and longer when they were not.
She opened the checklist. The first item was a source verification: did the attribution metadata match the originating system's registry-current status? She checked it. Yes. She moved to the second item.
Outside, the corridor was quiet. Somewhere across the city, Anitha was in the calibration floor or back at her desk and a message in her queue was waiting to be answered. Somewhere in the QA system, a three-sentence request was pending under a self-flagged input from a twenty-day provisional relay node. Somewhere in a committee's record, a question Kavya Sundaram had asked this morning was now traveling — to a clerk, to a docket, to whoever processed these things next.
She checked the second item and moved to the third.
The morning had been quiet. The afternoon would be work. What she had learned, over six months of inherited calibrations and portability-window adjustments and governance mandates, was this: the records moved without her when she let them move. That was not a failure of agency. That was the institution working. The question she kept returning to was whether she could keep working while they did. So far: yes. Most of the time. Getting better.