Two Doors
The attribution feed updated at 22:14, and Mitsue Hoshino read Kavya Sundaram's note twice before she moved.
She had been at her desk since 20:34, when Tomasz's message had arrived confirming the date: Wednesday, 14:00, clinical governance formal review, institutional representation required. He had used the word accountability, which had a legal weight she wasn't sure he'd intended or hadn't chosen deliberately. She had written it in her notebook — the paper one, not the document stack — and then closed the notebook, because a word with no subject attached to it didn't need more space yet.
The attribution feed was a narrower display, mounted flush to the left of her primary screen and angled slightly toward her. It pulled from the consortium's shared ledger: every lineage-interrupted flag logged in the past 72 hours, color-coded by routing status, scrolling slowly enough to read if she wasn't doing anything else. Most nights she ignored it. Tonight she had been watching it without quite looking at it, the way you watch a door.
The 22:14 update was flagged yellow, which meant it had been routed through the portability-transfer channel before arriving in the feed. Yellow meant the record existed. It did not mean the origin was clean.
She opened the note.
Four people at the Creative Workers Federation policy working group: two labor organizers, one contract lawyer, one researcher who studied platform governance. Kavya had sent the 20-case disclosure register as a resource, not a grievance, and the working group had read it overnight. The contract lawyer had made two phone calls and come back with a finding Kavya had not anticipated: the disclosure moment had no legal definition.
Kavya's four-condition standard — the one she'd been refining for the past two months, the one that was supposed to clarify when a client could be considered adequately informed that their aesthetic signature had been analyzed, stored, or transferred — assumed a specific moment in time when disclosure was legally required to happen. The contract lawyer had found that the moment did not exist. There was no legal architecture requiring the disclosure to occur at any particular point. Kavya's conditions were a clarification of something that had never been defined.
Mitsue read this twice. Then she looked up at the attribution feed.
The 11.9% lineage-interrupted rate in Synthesis-9's provisional relay node had been computed automatically. The lineage-interrupted flag existed because someone had designed it into the intake protocol. But the protocol had not been designed with a legal definition in mind; it had been designed with a practical one: if the provenance chain broke during portability-transfer, flag the record and route it for attribution review. The assumption was that attribution review would determine what to do next. The assumption was that attribution review was a channel, not an open question about whether anyone was required to use it.
She had been holding 94 lineage-interrupted records, 81 of them from the portability-transfer channel. The routing gap was not Mitsue's to name. But she understood, sitting at her desk on Saturday night while Berlin made its late-night sounds below the window — a tram, someone's audio bleeding through their earpiece as they passed — that Kavya's legal definition gap and her routing gap were not two different problems. They were the same machine viewed from two different rooms.
She did not write this down. It was too early for analogies, and it wasn't her work to say. She closed Kavya's note.
There were currently two paragraphs in PARAGRAPHS.
Bongani Sithole's had arrived first: The channel does not distinguish between research records and clinical records. That distinction was assumed to live in the review node. The review node was never built. He had left the second sentence without a subject. Mitsue had preserved it exactly — no subject supplied, no punctuation adjusted. The absence of a subject was doing work she didn't want to undo.
Anitha Venkataraman's had come twelve hours later: ninety-four records, eighty-one through the same door. She doesn't know if pattern or design. Anitha had used the third-person pronoun to describe Mitsue to Mitsue. Mitsue had read it three times before she accepted that it was a description and not an accusation. The she doesn't know was accurate. She still didn't know.
The submission pathway had been on hold since Tomasz's Friday night message: clinical governance director wants formal review, institutional representation required. Hold. The oversight board track and the clinical governance track were different legal structures. Oversight board review was advisory. Clinical governance formal review had enforcement capacity — not certain enforcement, but the structural capacity for it, which was what accountability meant in Tomasz's sentence.
The hold note in PARAGRAPHS read: Submission pathway under review — hold per Tomasz Krawczyk, 2026-08-07 23:47.
She had added nothing to it since Saturday morning, when she had written WAIT.txt: The question is not whether the data is right. The question is who has the authority to require a response.
That was still the question. Wednesday might answer the second half of it.
She opened a new file and named it PARAGRAPHS_CLINICAL.txt.
She pasted the PARAGRAPHS template header and then stopped to revise it. The standard oversight header read: Aggregate report, lineage-interrupted flags, Synthesis-9 session records, submitted to consortium oversight board per protocol 7.3. The clinical governance header needed to be different. She looked at the format Tomasz had forwarded from the director's office and typed: Prepared for Clinical Governance Formal Review, per director request, 2026-08-08.
Then she added a date field. Then she added an institutional affiliation field for each contributor.
She tabbed through the form to Bongani's entry. His paragraph was already there, pulled from PARAGRAPHS: The channel does not distinguish between research records and clinical records. That distinction was assumed to live in the review node. The review node was never built.
The institutional affiliation field was below his name. She stopped with the cursor in it.
She knew Bongani's institutional affiliation. She had looked it up in the consortium directory three weeks ago, when she was trying to understand whether his access to the portability-transfer channel records was through the same exemption as Priya's or a different one. Cardiology overflow, regional portability registry, listed under the distributed audit consortium as a secondary observer. It was not private information. She could type it in thirty seconds.
She moved the cursor to Anitha's institutional affiliation field instead.
Anitha's affiliation was also in the directory. Consortium primary observer, machine learning ethics subcommittee, formal appointment dated 2026-03-01. Anitha had been the one to name the Tuesday meeting a calibration census before anyone else used that word, which told Mitsue something about how Anitha understood the scope. It was not private information either.
She tabbed to Priya's slot. Priya's paragraph hadn't arrived yet. The slot held a placeholder: [PENDING — Priya Mehta].
Institutional affiliation: blank.
She tabbed to Tomasz's slot. [PENDING — Tomasz Krawczyk]. Institutional affiliation: blank.
She went back to Bongani's field. The cursor blinked in the empty institutional affiliation.
Bongani had not given it to her. He had contributed a paragraph to an oversight document. The oversight document had not required institutional affiliation because the oversight board protocol didn't ask for it. The clinical governance formal review template did ask for it. Bongani had not been told about the clinical governance track yet. He had not agreed to participate in it.
She left the field blank. She moved to Anitha's. Left it blank. Moved to the two pending slots. Left them blank.
The document now had four contributor entries: two with paragraphs and no institutional affiliation, two with placeholders and no institutional affiliation. A formal review header at the top. A date field, filled. An authority line — Clinical Governance Formal Review, director's office, 2026-08-08 — which implied a weight the data had not yet been asked to carry.
She would send each of them a note on Monday. She would explain both tracks — the oversight board submission, the clinical governance formal review — and ask which they were willing to participate in. She would include the formal review template so they could see what their name would appear on. She would not assume the answer was yes. They had contributed paragraphs to something she had asked for. The director's office was asking for something different.
That was a different question. It needed a different answer.
She saved PARAGRAPHS_CLINICAL.txt and kept it open beside PARAGRAPHS.
The two documents sat side by side on her primary screen. Same data. Different headers. The institutional affiliation fields in the clinical version all blank, waiting for Monday.
At 23:07, the attribution feed flagged a new entry, and Mitsue read it without moving from her chair.
Synthesis-9's registry review team had sent a specification request to the provisional relay node: they needed to classify the node's inputs as either first-hand observational data or processed interpretations of other observations. The distinction mattered for how the 11.9% lineage-interrupted rate would be weighted in the month-end registry review. The month-end registry review would determine whether the node's temporary exemption was sustainable. The exemption had 23 days.
Mitsue was not part of the registry review team. This was not her work.
She opened CHANNEL_NOTES.txt, scrolled to the bottom — below the two-pathway summary, below the enforcement capacity note, below the hold status update from Saturday afternoon — and added three lines:
Synthesis-9 registry review: data type classification requested (first-hand observational vs. processed interpretations). Adjacent question: if a lineage-interrupted flag marks a record that exists without traceable origin, is the flag itself first-hand data? Note for context only. Not in scope.
She closed the note. Not in scope was a position she had to keep choosing, every time a new adjacent question arrived in the feed. The machine had more rooms than the two she was in. Kavya was in one of them. Tomasz was in another. The registry review team was in a third. Mitsue was the one holding the paragraphs, which meant her job was to stay in the room with the paragraphs and not wander.
The attribution feed resumed its scroll. Three new lineage-interrupted entries, all from the portability-transfer channel, all tagged provisional pending review. Standard Saturday night cadence, she had learned over the past sixteen weeks: the system never stopped, it just processed more slowly on weekends, when the consortium's review staff was smaller.
By midnight, she had four documents open.
PARAGRAPHS: two entries, two placeholders. Bongani's paragraph about the review node that was never built; Anitha's 94 records, 81 through the same door, still not knowing if pattern or design. Two blank slots where Priya and Tomasz would eventually send their sentences.
PARAGRAPHS_CLINICAL.txt: the same data in a formal review header, every institutional affiliation field blank, waiting for Monday's notes and whoever said yes.
CHANNEL_NOTES.txt: two pathways, growing sideways annotations, the note about Synthesis-9 at the bottom.
WAIT.txt: two sentences, unrevised since Saturday morning. She had looked at them again around 23:30, half-thinking she might add something. She hadn't. The question is not whether the data is right. The question is who has the authority to require a response. Wednesday would give her the second half. Until then, it was still accurate.
She turned off the desk lamp.
The screens stayed on. The attribution feed continued its slow scroll in the dark, pale yellow flags drifting upward, each one a record that existed and whose origin could not be verified. She had stopped finding this disturbing sometime around the sixth week. Now she found it legible: the machine was doing what it had been designed to do, which was to flag the problem without resolving it. Resolution was a human function. The humans had not yet been assembled in the right room.
Wednesday, 14:00, the clinical governance formal review. Institutional representation required. The director had said accountability.
She would bring PARAGRAPHS. She would bring PARAGRAPHS_CLINICAL.txt, with its blank fields, and a note explaining why the fields were blank and what she would need before she could fill them. She would bring CHANNEL_NOTES.txt and the original six records from Synthesis-9 session 2847 and sixteen weeks of evidence that the flag had not been a mistake and had not been smoothed and had not been made easier to look at.
Two doors. She had prepared for both.
Outside the window, the building across the street had gone dark on the fourth floor. The tram had stopped running for the night. The attribution feed scrolled without commentary. Kavya was probably still working through the legal definition gap in whatever notes she kept. The registry review team would send its classification request to the provisional relay node again on Monday, if the node's operator didn't respond over the weekend.
Mitsue went to make tea.
She did not know which door would open or when. That was exactly the architecture: two pathways, two document versions, four blank institutional affiliation fields, and a WAIT.txt that was still accurate.
The water came up to heat in the dark of the kitchen. The attribution feed's pale glow reached through the doorway from her desk.
She stood there and let it.