The word was "decisional."
Mitsue Hoshino read Tomasz's message at 07:14 on Sunday morning standing at the kitchen counter with her first cup of coffee, still in the clothes she had slept in. The ambient display above the counter showed the building at 31% of peak energy — lower than yesterday, quieter Sunday. In the bottom-right corner of the display, a soft status bar showed the Synthesis-9 coordination load at 12%, light blue, well within normal range for a weekend batch cycle. Through the window the Prenzlauer Berg morning was already warm, pigeons on the ledge, the first pedestrians, the particular quality of Sunday light that is unhurried in a way weekday light is not.
She read the message twice.
Clinical governance director's office has confirmed: Monday's meeting is decisional. She will attend in person. Please review any materials you plan to present — the meeting will produce binding guidance.
Decisional was a specific term in the consortium's governance framework. She knew this because she had read the governance documentation in May, looking for the mechanism behind exemption 7(b), and had read more than she had intended to. Decisional meant the meeting would end with a resolution. Not a recommendation. Not a preliminary assessment. Not a note in the record that additional review was warranted. A resolution — binding, institutionally enforceable, going into the compliance record and not advisory.
The clinical governance director had enforcement authority the oversight board did not have. Mitsue had written that in CHANNEL_NOTES.txt. She had written it as a distinction between two channels, two levels of consequence, when she was trying to understand which channel her PARAGRAPHS document should travel through. She had thought Tomasz was handling the channel question. She had thought she was preparing materials for Tomasz to deliver. She had prepared ONE_PAGE_GOVERNANCE.txt as a record transfer, 201 words, with a final line she had been careful about: This summary is not a recommendation. It is a record transfer.
She put her coffee down and opened her laptop.
ONE_PAGE_GOVERNANCE.txt was still 201 words. She read it from the beginning.
Two accurate numbers: 5.1% across all portability-transfers, 11.9% concentrated in the provisional relay node. The same 94 attribution-contested records, counted two different ways. Attribution-contested was the term of art for records in the queue: outputs from the synthesis network that had generated a lineage-interrupted flag during portability-transfer review and had not yet been marked attribution-clear. The distinction mattered. Attribution-clear meant a human reviewer had examined the flag, traced the lineage, and certified the output. Attribution-contested meant the flag existed and the certification did not. Her initials were in the reviewer field on all 94. The records were not unresolved because she had failed to review them. They were unresolved because resolving them required governance authority she did not have.
The exemption 7(b) explanation in three sentences: the attestation service had checked the lineage, found the source registry provisional-unverified, applied the exemption, forwarded the output, and generated the lineage-interrupted flag. The flag was the system working as designed. Then the 23-day countdown: provisional node exemption expires month-end. Then the available-on-request list — CHANNEL_NOTES.txt, the archive export of 94 records, the attestation service logs, the two-population analysis (83 skip-step, 11 query-delay), scope statements v1 through v4, the PARAGRAPHS document still on hold.
Final line: Note: this summary is not a recommendation. It is a record transfer.
She had written that sentence on Saturday afternoon. She had written it because she had finally understood what she was doing — she had been writing flags, not briefs. The lineage-interrupted flag from the attestation service said: here is what I found, here is the number, a human should review this. Her ONE_PAGE_GOVERNANCE.txt said: here is what I found, here are the numbers, the governance committee should review this. Same structure, different scale. She had felt that she had gotten something right.
Now she was looking at the sentence and feeling that it might be incomplete.
If Monday's meeting was decisional — if the clinical governance director was attending with binding authority — then the meeting would produce guidance. Guidance about what? The provisional relay node's exemption status, presumably. Renewal, termination, conversion to verified status. Those were the obvious outcomes of a governance review of the node.
None of those outcomes would mark 94 attribution-contested records attribution-clear.
She had written in CHANNEL_NOTES.txt that the records and the node were separate timelines. She had written it as documentation, as a note to be available on request. She had not written it as a recommendation because that was not her function. The auditor's function was to flag, escalate, document, and let the channel decide. She had done all of those things.
But she knew things about these records that were not in a 201-word summary. She had been working with them for six months. She knew the two-population structure — 83 skip-step and 11 query-delay — and what it implied about two different failure modes inside one flag category. She knew the 11.9% vs 5.1% framing and why it mattered for institutional decision-making. She knew the exemption 7(b) mechanism well enough to explain it in plain language to a room of clinical administrators. And she knew what resolution should look like. She had known for a while and had been not-saying it because not-saying it was the right posture for the work she had been doing.
A decisional meeting was a different kind of room.
She put her fingers on the keyboard.
She typed: Recommendation: extend provisional relay node exemption for 30 additional days to enable resolution of the 94 attribution-contested records.
Read it back. Deleted it. The problem was that it was about the node. She did not have a strong view about what the node should do. She had a view about what needed to happen to the records.
She typed: The 94 attribution-contested records currently in the queue require dedicated governance action independent of the node exemption decision.
Better. More accurate. But still too flat — it did not say what the action should look like, or who should be responsible, or what the timeline should be.
She typed a third version: Note from attribution trace clerk, added 2026-08-10: The 94 records currently in the attribution-contested queue require direct governance action independent of any decision about the provisional relay node's exemption status. These records predate this governance review. They will remain attribution-contested after it unless resolution is explicitly included in Monday's binding guidance. A specific timeline and responsible party for record resolution should be part of the committee's decision. The attribution trace clerk is available to support the resolution process. This note supplements the record transfer above and represents the clerk's assessment, not a formal institutional recommendation.
She read it back carefully. It was 97 words. It said what she thought. It named the gap between a node-exemption decision (which would govern new transfers) and the existing 94 records (which needed direct action regardless of what the node did). It labeled itself clearly as the clerk's assessment, not a recommendation, because she did not have the institutional standing to make a recommendation in a governance proceeding — but she had the standing to say, clearly, what she had observed and what she thought needed to happen. That was the flagging function. That was the auditor's job. She had been thinking the auditor's job ended at "a human should review this." But there were humans in the room now, and the review was happening Monday, and if she did not say what she thought the review needed to include, she would be withholding something she knew.
She saved the file. 298 words. The last 97 were hers.
At 08:02 the Sunday batch notification from Synthesis-9 arrived in her attribution feed with an amber alert indicator — unusual for a routine summary.
She opened it.
Week 38 Sunday batch: 312 inputs processed. Input 1,797 from provisional relay node: confidence interval 0.81, below the 0.83 baseline established in weeks 35–37. Synthesis-9 querying registry team for calibration guidance. Human review of input 1,797 requested.
She read this twice, then a third time.
Synthesis-9 had flagged its own output. The protocol for this was different from the attestation service's lineage-interrupted flag. When the attestation service generated a lineage-interrupted flag, it was reporting an external condition: the transfer chain had a gap, and here was the location of the gap. The flag was about the infrastructure. But when Synthesis-9 flagged its own output, it was reporting an internal condition: one of its outputs had fallen below the confidence baseline it had established for itself over the preceding three weeks. The flag was about Synthesis-9's own synthesis quality. Two different failure modes, two different flagging mechanisms, both landing in the same attribution feed on the same Sunday morning.
The amber alert meant human review was requested, not required. The protocol said: a synthesis quality flag at confidence 0.81 (two percentage points below baseline) should be reviewed by the registry team within 72 hours. It did not generate a lineage-interrupted flag on the portability-transfer layer unless the record actually went through a portability-transfer — and she did not know yet whether input 1,797 would. The amber alert would sit in the feed, visible to anyone monitoring it, until the registry team acknowledged it or 72 hours passed and it escalated automatically to the next tier.
She was not on duty. She had decided Sunday was not for the attribution gap.
But she noted it. She wrote a line in the attribution record: 2026-08-10 08:02 — Synthesis-9 input 1797, confidence 0.81 (below 0.83 baseline), querying registry team, human review requested. Not reviewed today. Will assess post-governance call Monday.
She closed the attribution feed. The status bar above the counter had not changed — Synthesis-9 coordination load still at 12%, light blue.
She finished the coffee. It had gone warm rather than cold — the Sunday morning pace, less urgency. Through the window the Prenzlauer Berg market was beginning to set up: the first van, the first folding tables, someone carrying a crate. A small routing node on the building's exterior wall blinked amber for three seconds — probably a neighbor's morning AI sync, nothing unusual — then returned to steady green. She had been sitting at this table since 07:14. It was now 08:15.
She had told the governance director's meeting what she thought the binding decision needed to include. She had said it clearly, labeled it clearly as her assessment rather than a recommendation, and preserved the record-transfer framing for the main body of the document. She had not been prescriptive about the node's exemption status because she genuinely did not know what the right answer was on that question. She had been prescriptive about the 94 attribution-contested records because she had been working with them for six months and she knew they needed their own resolution timeline and she was the one who knew it.
The provisional relay node's exemption had twenty-two days remaining.
Synthesis-9's input 1,797 was in the attribution feed waiting for registry review.
The 94 attribution-contested records had her initials in the reviewer field, and now there was a sentence in a governance document that said they required their own resolution timeline and a responsible party. That sentence had her name on it. It was labeled clerk's assessment. It was not binding. It was evidence that someone who had read every one of those records thought the records needed to be in the room on Monday.
She put on her shoes and went outside.
Sunday in Prenzlauer Berg was what it was: warm, green, the market setting up, the smell of coffee from the bakery on the corner, a child on a bicycle too large for them negotiating the curb. At the corner of Kastanienallee, a small infrastructure panel displayed the neighborhood's current AI coordination load — 8% above baseline for a Sunday, the display said, attributed to an automated batch run scheduled without manual override. Someone had forgotten to set the off-peak window. The panel would log a soft inefficiency flag and clear itself. Nobody would review it. Nobody needed to. That was the point.
The portability-transfer queue was running its Sunday batch in the consortium's server rooms without her. Synthesis-9 was processing. The governance director's office was preparing for Monday.
She had said what she thought. It had taken twelve minutes to say it, after six months of not saying it. She did not know if it was the right thing to say, in the way that she often did not know whether a flag she had filed would be validated by subsequent review or quietly set aside. She had filed the flag. She had put her name on it. She had said: this is my assessment, take it for what it is.
She walked out into the August morning and let the door close behind her.
She had thought, when she started in April, that the function of the attribution trace clerk was to maintain the integrity of the lineage record — to ensure that every output from the synthesis network could be traced back through the portability-transfer chain to its verified source. That was the job as she had understood it. Flag the gaps. Contest the attributions. Keep the record attribution-clean.
What she understood now was that the record's cleanliness depended on the humans who reviewed the flags. The attestation service generated the lineage-interrupted flag. She reviewed it. The governance committee would receive the summary. The clinical governance director would make the binding decision. Each of those steps required a human being to look at a record and decide what it meant and what should happen next. The record could be perfectly maintained — every flag filed correctly, every attribution contested appropriately, every exemption application documented — and still be attribution-contested, because marking a record attribution-clear required the right human being to be in the right room with the right authority.
She had been trying to make sure the right information would be in the room. That was what the past three weeks had been: not case-building, not advocacy, but information stewardship. Making sure that when the decisional meeting happened, everything the governance director needed would be on the table.
Including what Mitsue thought.
The market smelled like cut flowers and warm bread. She did not know if the 97 words were enough. She did not know if the governance director would read the clerk's assessment at the bottom of a 298-word document or skip to the numbers. She knew the 94 records. She had put what she knew into writing. The rest was not hers to carry.
She bought a coffee from the bakery and stood with it in the early warmth, watching the market finish setting up, the attribution feed quiet in her pocket, Synthesis-9's amber alert logged and waiting, the week beginning.