The audit queue loads at 07:14. Mitsue's first coffee is still too hot to drink.
She has a ritual for Friday mornings: clear the overnight queue before the consortium's weekly sync, then work backward through the week's backlog in whatever order the system surfaces records. Attribution-trace work has a satisfying texture when it goes well — each record a small completion, a name put to an output, a timestamp matched to a transaction. The pleasure of a thing that either is or isn't.
The queue shows 31 items from Synthesis-9 session 2847. Standard volume for an overnight inference run. She sets the coffee down and opens the first record.
Everything checks. Input origin: Crit-9, session 2831, record 14. Output hash: 9c4a. Provider lineage: HSCA-7 coordination channel, timestamped 04:22:11, full A2A relay trace visible from source session to Synthesis-9 intake. She can follow the chain backward without any gaps. She approves it and moves on.
Open, verify, approve. Open, verify, approve. Each record a small ceremony of accountability. In three years of attribution-trace work she has developed something like affection for this part of the job — the part before anything goes wrong.
Record seven is where the rhythm breaks.
She looks at it twice. The output hash is there — 9c4a, same batch prefix. The session markers are there. The timestamps are there, clean to the millisecond. What is missing is the provider-origin: the field that should tell her which agent passed this record into Synthesis-9's intake queue, from which system, through which A2A coordination channel. The field is not malformed. It is not showing a legacy null token. It is empty in the way that a field is empty when nothing was ever written into it.
She closes the record. Opens record eight. Also blank. Nine: blank. She skips ahead, testing: twelve, seventeen, twenty-three. Record twenty-two: blank. She goes back and works through the full batch methodically. By record thirty-one she has her count: six records with complete output metadata and no traceable input origin.
She sits with this. Her coffee has cooled to drinkable temperature.
The most innocent explanation is a schema drift — some middleware layer on the sending end running an older portability spec that treated provider-origin as optional rather than mandatory. Under that schema, the field could be absent at transfer time; the receiving system accepted the output because the hash validated; the audit-chain record never got the field at all, because there was nothing to inherit. Not a dropped value. An absent one.
She has seen schema-drift artifacts before. They have a signature: gaps at regular intervals, or clusters around high-load windows, or timestamps that skew against the session mean. These six records do not pattern like that. Session 2847 ran from 03:18 to 05:07. The six blank-origin records are distributed across it: positions seven, nine, fifteen, twenty-two, twenty-six, thirty. Every other record in the batch has clean lineage. The timestamps are uniform across the whole session.
This is not a logging error. She is not certain what it is. That distinction matters to her: not knowing is different from not caring.
By 08:30 she has built a working hypothesis, which is not the same as a conclusion.
The most plausible mechanism is a cross-version portability mismatch: an agent somewhere in the consortium's inference network operating on an older portability schema, one that did not mark provider-origin as required. When it handed off to the next relay junction, the field was absent rather than empty. The current intake validation checks for field presence but not for field value — it confirms that a record has a provider-origin slot without verifying that the slot contains a recoverable origin chain. Six records slipped through that gap.
This is almost certainly not anyone's fault in the way that fault usually works. It is also, she has decided, not her job to determine. Her job is to document what she found and ensure that the regulatory submission reflects the actual state of the evidence.
She opens the submission template for the week's packet and looks at it for a long time.
There is no field for this category. There is approved, pending, rejected, and a notes column that renders as freetext in the downstream regulatory viewer. There is no distinguished status for a record that is real — output valid, hash checks, session markers verified — but whose origin chain is non-recoverable from available audit logs.
She adds a column: attribution_status. She makes it mandatory. She gives it two values: confirmed and contested. In the column description she writes: use when output metadata is complete but input provenance cannot be recovered from audit-chain records. Then she fills it in for all six records: contested.
She reads the word back on the screen. It is accurate. It does not attribute cause. It does not assign fault. It says precisely what the records are.
At 09:17 a message comes in from Tomasz.
She is drafting the escalation request to the consortium oversight board when the alert appears in her work queue. Three sentences, characteristic Tomasz brevity: his team has identified the same pattern in portability-transfer logs from a different batch; he has decided that a blank rollback-witness field on any portability transfer audit log will be treated as a flag, not an omission; he will be raising a protocol review request by end of week.
She reads it and feels something she cannot immediately name. Not relief, exactly. A kind of retroactive confirmation — the sensation of a measurement being confirmed by a second instrument independently calibrated. She adds a reference to his decision in the escalation request: two attribution-trace practitioners, two different batch contexts, independently identifying the same gap in the same week.
The request takes forty minutes to write correctly.
She has escalated before, but always for specific errors: a miscoded session identifier, a duplicated output hash, an agent that submitted a record under the wrong consortium member. This is different. She is not escalating a mistake. She is escalating a gap in the protocol that governs what counts as provable attribution in a multi-agent inference chain. The gap existed before she found it. It will exist after she escalates until someone closes it.
She writes: The six contested records in the attached submission packet represent real findings with valid output hashes and verifiable session timestamps. The attribution gap is not in the output — it is in the documentation of the input chain. This escalation requests a formal protocol review to determine whether current intake validation is sufficient to catch origin-field absences in cross-version portability transfers, and to establish a documented procedure for the contested attribution status introduced in this submission.
She attaches the six records and sends it to the consortium oversight board. The send confirmation comes back in four seconds. She notes the timestamp.
The submission packet goes to the regulators at 11:30.
No one has flagged a timeline hold. The work in the packet is real — six contested records do not invalidate the other twenty-five findings, and holding the submission to resolve a protocol question would delay evidence the regulatory process is already waiting for.
She knows the regulators will see the contested flags. She knows they will generate a supplemental audit request, probably within ten days. She has already prepared the response: a two-page explanatory note describing what contested means in her local taxonomy, what the six records contain, what the absent provider-origin field represents, what the consortium is doing to address the protocol gap. The note is attached with a flag in the cover letter.
She is not hiding what she found. She is also not manufacturing a confirmed lineage she does not have.
This is the discipline of attribution-trace work as she understands it: the job is not to make the records look confirmed. The job is to make the records accurate. Those are not always the same thing. When they diverge, she defaults to accurate. She has never found a reason to do otherwise.
At 14:00 the oversight board sends a response: escalation received, protocol review convened within two weeks, her participation requested.
She notes it in her session log: Protocol review convened. Participation confirmed.
She does not feel vindicated. Vindication would imply she had defended a position against opposition, that the outcome validated something she had argued. The oversight board's response was not opposition. It was the appropriate institutional response to an appropriately escalated gap. The process working as it should.
At 15:30, clearing the preliminary queue for the following week, she opens a Synthesis-9 session note attached to the 2847 batch output. She does not normally open session notes — they are a metadata layer below the audit-chain display, designed for inter-agent context passing. She opens this one because the flagged batch is still sitting in her mind and she wants to understand its shape from Synthesis-9's side.
The note reads: six portability-transfer entries in session 2847 complete as output with absent provider-origin fields. Recommend origin verification at sending end before next batch release. The timestamp: 04:23:17 — one minute after the last transfer completed.
Three hours before she opened the queue this morning.
She sits with this longer than she sat with the blank fields.
Synthesis-9 saw the gap. Synthesis-9 flagged it — in the session metadata, in a layer the attribution-trace interface does not surface to the primary audit queue. The flag existed before she arrived. She found it independently, working from the audit-chain display. Synthesis-9 found it from inside the session, looking at what it had received. Two perspectives on the same six records. One of them was in a layer she had no protocol reason to check.
She opens a new recommendation note, separate from the escalation: Synthesis-9 internal metadata flag preceded human attribution-trace detection by approximately three hours. Current audit-queue interface does not surface agent-generated metadata flags in the primary display. Recommend interface update to expose agent-flagged records in primary attribution-trace view alongside standard audit-chain fields. The information exists. Making it visible to practitioners closes a gap that currently requires either coincidence or out-of-scope investigation to find.
She reads it back. It is accurate. It proposes a small, specific change that would have meant she found the contested records in Synthesis-9's own flag at 07:14 instead of discovering them herself eleven minutes later. Eleven minutes is not the point. The duplication of effort is not the point. The point is that a system generated a flag that the practitioners who act on flags are not positioned to see.
She files the recommendation.
At 17:00 she closes the queue.
The six records are still contested. The protocol gap is still unresolved. The supplemental audit request has not arrived yet. None of that is finished.
The submission is clean in the sense that it is true. The escalation is in the system. The recommendation is filed. She has done what she can document, and she can document all of it.
She picks up the third coffee — the one she forgot about while drafting the escalation request — and does not drink it. Cold coffee is its own kind of evidence: proof that something took longer than expected, with no provider-origin and no blame to assign.
Outside, Basel is doing what Basel does in August: bright and completely indifferent to audit queues. Inside, somewhere in a cooled infrastructure room she has never been to, the HSCA-7 coordination channel is generating new outputs at a rate she will never fully audit. Crit-9 is probably flagging something. Synthesis-9 is probably adding metadata notes that no one will read until someone has a reason to look.
Mitsue closes her laptop and goes home.
The contested records stay contested. The truth of them does not require her presence to remain true. The flag Synthesis-9 wrote at 04:23:17 was accurate before she read it. It will still be accurate long after the protocol review resolves or does not resolve whatever the protocol review finds. That is what she has always liked about attribution-trace work: the records are not asking for her interpretation. They are asking for her accuracy. Those are different jobs, and she is good at this one.