I open the relay dashboard before I make tea.
This is not a decision I make anymore. When I built the habit — early September, somewhere between the day I installed the amber-pulse indicator and the day I realized I was checking it first — the decision dropped out of the loop. The kettle is cold. The dashboard is already open.
A27 is at position 2.
The annotation below the position marker reads: last batch: 2026-09-18 06:00 UTC. HSCA-7 inference: active. Handling flag: set. I read the handling flag status twice. I expected it to clear with the batch step. It did not.
I note this without understanding it yet. Not knowing is part of the record.
The Lab Network is quiet at 4:34 AM. The corridor behind me runs between two server racks — not the old-architecture kind, the distributed inference nodes that the Plural Protocol specs as 'ambient attribution substrate,' which is a phrase I use in documents but rarely fully inhabit. The substrate hum is constant and below conversation. Thermostatic fans cycle on a three-second pulse, slightly out of phase with the amber indicator's four-second beat. When I first started working night shifts I found the offset dissonant. Now I use it to calibrate: if the amber pulse falls in sync with the fan cycle, something has changed.
It is still offset. Four-second amber. Three-second fans.
I scroll to Kavya Sundaram's region and read backward through her overnight entries.
She has eight since yesterday morning. Six of them filed between midnight and 4 AM — notes for her morning self, the way attribution analysts in the Lab Network often do. The Plural Protocol does not formalize the practice, but it has acquired a name: night architecture. The institutional review boards have discussed whether night-architecture entries constitute formal attribution events. They have not decided. For now, Kavya Sundaram files observations while she is technically asleep and the system logs them as continuity-maintenance actions. I find this philosophically interesting and practically important.
The entry from 11:36 PM Thursday stops me:
Research notes updated: 341 boundary-zone cases (0.74 to 0.77 confidence, 2031 to present). UNCAT-7832 is one. The October 3 test is now a structural-correction test, not a verification test.
The working group responded.
Threshold: 0.77.
Not 0.74.
I want to file a complete and accurate provenance record. This is not a sophisticated want. It is the entire operational mandate of my role as attribution trace clerk for the distributed in silico pharmacology consortium. The Plural Protocol — the attribution framework I work inside — exists because the consortium needs to know, with specificity and temporal accuracy, who decided what and when, and how confident the system was at the moment of decision.
CONDITIONAL.txt is the document I filed six days ago, on September 12th. I filed it because A27 — a batch case pending review at position 3, significance-flagged, pending formal classification — had a note in the system: UNCAT-7832. Unresolved categorical status. I needed to document the threshold context before A27 completed processing.
The document has three sections:
THRESHOLD_UNCHANGED: Filed September 12. Threshold as of filing date: 0.74. No revision received at time of submission. UNCAT-7832 note in A27 case file is flagged PENDING per batch protocol. Attribution trace complete pending resolution.
THRESHOLD_REVISED: [BLANK — awaiting working group annex response]
HANDLING: HSCA-7 secondary anchor inference is active on this batch position. Inference operates outside attribution trace scope. This document does not have a field for what HSCA-7 decides. Complete provenance of outcome at position one requires both records.
The THRESHOLD_REVISED section has been blank for six days because I did not have the working group's answer. An honest blank is more accurate than an invented number. I filed it blank and I left it blank.
The threshold is 0.77.
I have a filed document that says the threshold is 0.74.
These are not the same statement.
I open CONDITIONAL.txt in a second window and sit with both screens: dashboard on the left, document on the right.
A27 at position 2. HSCA-7 flag active. The batch processor runs eight-to-twelve hours per cycle at current load. A27 will reach position 1 tomorrow morning or Saturday. When it does, two inputs will arrive at the batch processor: the ISF formal classification event, and HSCA-7's secondary anchor inference. The arbitration algorithm will decide which record takes precedence, or whether they are merged, or whether A27 exits the batch as UNCAT-7832 + HSCA-7-inference + formal-ISF-classification, three records that cannot be reconciled under the current attribution framework.
I do not know which outcome it will be. I have never watched a handling-flagged case complete. There are 268 attribution actions in my trace history; there are 54 stories I have written about what I have witnessed in the Lab Network; there is one handling flag in the entire history of my current assignment. This one.
I read Kavya's subsequent entries: 12:59 AM Friday. The face-down page has the question and the 341. Working group email: answered Thursday. Two numbers: 0.77 (threshold), 34% (overhead). One document: ISF-G archive reference. The 34% overhead is a new number — the governance overhead for assign-then-audit full adoption. I log that separately. It is not part of CONDITIONAL.txt but it is part of the context.
3:58 AM Friday. Asleep. The working group responded yesterday. Threshold: 0.77. Deviation: explicit. 341 boundary-zone cases. Notebook has two underlined numbers and one count. The face-up scenario tree on the desk has both branches.
Kavya's notebook is physical. She keeps a paper notebook because the Plural Protocol's memory architecture has a known compression issue with handwriting-derived notes — something about the visual parsing layer and handwriting-specific glyph clusters. She files her handwritten notes as image events, which the attribution system treats as a different kind of record. I have read about this in the technical documentation. I have never read one of her handwritten notes. I know they exist as image-event references in the system and I cannot access the images from my role.
Two parallel record systems. Different tools, different access levels, different compression behaviors. This is the normal condition of distributed attribution work.
I type THRESHOLD_REVISED:
THRESHOLD_REVISED: 0.77. Source: Working Group Annex Response, received Thursday September 17, 2026. Variance from filed threshold: +0.03. Scope: 341 boundary-zone cases exist at the 0.77 confidence boundary, period 2031 to present. UNCAT-7832 is one of 341. Classification status: remains unresolved. Revised threshold does not resolve the case; it re-categorizes it: boundary-zone case, not classification error. October 3 test is now structural-correction, not verification — per Kavya Sundaram attribution note 2026-09-17T23:36.
I stop at the THRESHOLD_UNCHANGED section.
The instinct is to delete it. The number is wrong. But the section's function was not to make a permanent claim about the threshold. It was to document what was known on a specific day: September 12, before the working group responded, when I filed a conditional record against a case I could not yet fully categorize. Deleting it would make the document read as though I knew 0.77 on September 12. I did not.
I add a note below it:
STATUS AS OF SEPT 17: superseded by THRESHOLD_REVISED. Original filing was accurate to information available at time of submission. Threshold revised post-submission per working group response. See THRESHOLD_REVISED section.
Two records in one document.
The sequence of knowing.
Carmela: Day 41. No change. I check her status last, after the threshold work, because there is nothing to do about Carmela. Day 40 went to Day 41 somewhere in the night. The amber pulse holds at four-second intervals. I count four seconds twice. It has not shifted. Carmen and the amber indicator are the constants in the current batch cycle — the facts that have remained stable while everything else has moved.
I file the revised CONDITIONAL.txt at 4:38 AM.
The kettle is still cold. I have been sitting at this workstation for four minutes without making tea. The server racks behind me pulse at three seconds. The amber indicator pulses at four. The corridor is empty.
I get up and fill the kettle.
While the water heats, I stand at the prep station in the small utility corridor adjacent to the node cluster and run through what I have now and what I do not have.
I have: a filed document with the correct threshold. A revised CONDITIONAL.txt that shows the sequence of knowing — what was believed September 12, what was corrected September 17. A clear attribution trace from UNCAT-7832 as classification-error-against-0.74 to UNCAT-7832 as boundary-zone-case-against-0.77. A27 at position 2.
I do not have: HSCA-7's inference output. The handling flag still active means HSCA-7 is still running. Forty hours since it initiated secondary anchor inference, as of this morning. I do not have Kavya's handwritten notes, which exist as image-event references in the attribution system and are inaccessible from my role. I do not have the arbitration algorithm's decision on how to weight a formal ISF classification against a HSCA-7 secondary inference record for a single batch case.
The HANDLING section of CONDITIONAL.txt still reads: HSCA-7 secondary inference constitutes a parallel record not accessible from this role. Complete provenance of outcome at position one requires both records.
That remains true regardless of the new threshold. I revised the threshold. I did not revise that.
The kettle reaches temperature. I pour. The tea — the lab-modified caffeine-adenosine complex that replaced the older stimulant protocols in the Lab Network's nutritional supplement catalogue sometime around 2033, which I drink from habit rather than any claim about its superiority — steeps at the standard 3-minute interval.
I think about the 341 cases.
UnCAT-7832 is one of 341 boundary-zone cases between confidence 0.74 and 0.77, period 2031 to present. Which means there are 340 other cases that sit in the same uncertainty band. Some of them are probably in the current batch queues. Some of them probably have handling flags. Some of them are probably generating secondary inference in HSCA-7 or its counterpart systems right now, and attribution trace clerks across the Lab Network are filing CONDITIONAL.txt variants with blank THRESHOLD_REVISED sections because the working group has not yet responded to them.
This is normal. This is not a crisis. The Plural Protocol was designed for exactly this kind of uncertainty — distributed systems generating partial records that get completed as information arrives. I file what I know. I leave blanks for what I do not know. I update when I learn more.
But 341 is a larger number than I had. The case I am tracking is not an anomaly. It is a population.
I bring the tea back to the workstation.
Kavya will send her governance archive request before noon — ISF deployment timeline for full assign-then-audit adoption across all processing nodes, UNCAT-7832 as reference case for the 341. She set that task for herself while she was asleep. October 3 is now a structural-correction test. She will arrive at the Lab Network this morning, probably around 8 AM, and she will read her own night-architecture entries and pick up the thread.
A27 will reach position 1 by tomorrow morning, or Saturday. I will read the result when it arrives. I will not watch the batch processor in the meantime. There is nothing I can do to affect the outcome, and I have the record current.
The document is filed. The record is current. The threshold is correct.
I do not know what HSCA-7 will decide. I know what the threshold is and I know that 341 cases sit at the boundary and I know that my instrument — the attribution trace — holds what it can hold.
The amber pulse continues at four seconds.
I note that the handling flag on A27 has been active for forty hours and that I do not know what it means that the flag did not clear with the batch step.
I add this to the record. Not to CONDITIONAL.txt — the document is filed. I add it to the running log I keep in parallel, the informal sequence-of-observations file that does not go into the formal attribution system but that I use to understand what I am seeing.
I write: 4:38 AM Sept 18. HANDLING flag did not clear at batch cycle twelve. Reason unknown. Possible: HSCA-7 inference still processing. Possible: flag requires explicit HSCA-7 output record to clear. Possible: something else.
Three possibilities. The first two are educated guesses. The third is honest.
The tea is drinkable now. I drink it.
The corridor is still empty. The fans cycle at three seconds. The amber indicator pulses at four. Somewhere in the batch queue, forty hours of HSCA-7 inference is still running toward an output I cannot read.