THE THRESHOLD WAS RAISED
The working group's email arrived at 2:55 PM. At 3:05, Kavya opened it.
She had a rule for this kind of email. She had developed it in her second year with the ISF analysis project, when she noticed that reading an answer in the moment of recognizing the subject line contaminated the first read with the noise of relief or disappointment before the content had time to be processed on its own. So she waited. At 3:00 PM she saw the subject — Re: Annex Request KAVYA-ISF-2026-09-15 — and closed the inbox. For ten minutes she sat with the scenario tree she had been building since Tuesday: if the threshold was still 0.74, UNCAT-7832 was above it and should have been assigned. If the threshold was raised, the boundary had shifted and the question changed. If the deviation from flag-before-assign was documented, the 2031 implementation had made a deliberate choice. If it was undocumented, the deviation was either error or unreported judgment. She ran through the branches cleanly. Then she opened the email.
The working group's message was two paragraphs. No preamble, which she appreciated.
The first paragraph addressed the anchor_confidence_threshold. The threshold as of 2026 was 0.77. It had been 0.74 at the 2028 working group specification — she had confirmed this from the original document, so this was not a correction, just a confirmation. The adjustment from 0.74 to 0.77 had been made in the 2031 implementation cycle. The working group's note said: "The boundary adjustment reflected volume scaling requirements that were not anticipated in the 2028 specification." No further elaboration. She didn't need any. She had read the 2031 processing capacity paper in August.
She wrote in her notebook: 0.77. Underlined once.
The second paragraph addressed the flag-before-assign deviation. The 2031 implementation had deviated from the 2028 protocol. The deviation was explicit and documented, not inadvertent. The working group quoted from the implementation decision record, document ISF-GOV-2031-IMPL-047: "Assign-then-audit was adopted explicitly to address processing velocity above 10,000 cases per cycle. The flag-before-assign model, tested in 2030, introduced 34% processing overhead at scale, which exceeded acceptable latency thresholds for the ISF batch cycle." The working group added: this decision record has been publicly available in the governance archive since 2032.
She wrote: 34% overhead. Underlined once.
She closed the email.
She had two answers. She now needed to do one calculation.
The ISF extraction record for UNCAT-7832 showed a style confidence score of 0.75. She had requested this record in August, knowing she would need the raw score eventually, not knowing yet what threshold she would need to compare it against. The record had arrived in the third week of August. She had filed it without acting on it, waiting.
Under the 0.74 threshold: 0.75 is above the boundary. The case is assignable. Under the 0.74 specification, UNCAT-7832 would have been flag-before-assign — flagged as a boundary case, reviewed before assignment. Under the flag-before-assign model, the review was designed to catch cases that numerical confidence alone might misclassify.
Under the 0.77 threshold: 0.75 is below the boundary. The case is below the cutoff. Under the 2028 model, a below-threshold case would be flagged for human review before any assignment decision. Under the 2031 assign-then-audit model, a below-threshold case would be assigned anyway — based on the best available anchor match — and then reviewed in the post-assignment audit cycle.
UNCAT-7832's outcome in October 2041 was UNCAT. Not assigned. Not flagged for review. Not in the audit queue. Just UNCAT: no anchor matched, no assignment made.
This was the anomaly.
Under the 2028 flag-before-assign protocol, a case with a confidence score of 0.75 below the 0.74 threshold would have been flagged for human review. A reviewed case might end up assigned or legitimately UNCAT, but the UNCAT designation would come after human involvement, not instead of it.
Under the 2031 assign-then-audit protocol, a case below the 0.77 threshold would be assigned — because the point of assign-then-audit was to assign everything and review afterward, not to pre-screen. The UNCAT outcome should not exist under assign-then-audit at all, for cases that match any anchor above some minimum threshold. A pure UNCAT, with no assignment and no audit trail, was a pre-2031 processing pattern.
She looked at the UNCAT-7832 processing record again. October 2041. UNCAT. No assignment. No audit note. No flag record. The record had the UNCAT outcome and nothing else.
She wrote: UNCAT-7832 scored 0.75. 2041 threshold was 0.77. Under assign-then-audit (2031+), 0.75 should be assigned + audited, not UNCAT. UNCAT in 2041 is anomalous.
She sat with that sentence for a moment.
The ISF batch processing cycle had two components: the classification engine and the assignment engine. The classification engine ran on the confidence threshold model. The assignment engine handled what happened to cases below threshold: pre-2031, they were flagged; post-2031, they were assigned. The UNCAT outcome was what happened when neither engine made a disposition — when a case fell below some minimum match score across all available anchors, or when the classification engine and assignment engine disagreed in a particular way. UNCAT was not a routine below-threshold outcome. It was an exceptional one.
What she did not know yet was whether the 2041 instance of the ISF system was fully running the 2031 protocol. Implementation dates were frequently not the same as full-adoption dates. If the assign-then-audit model had been adopted as policy in 2031 but not fully deployed in all processing nodes until, say, 2033 or 2035, then cases processed in 2041 might still be running on legacy logic in some part of the system. UNCAT-7832 might have been processed by a node that had not yet completed the migration.
This was a separate research question from the one she had asked the working group. She noted it in her notebook: check ISF deployment timeline for assign-then-audit full adoption across all nodes.
She then opened the ISF public statistics interface and ran a search she had been holding in reserve. She filtered for cases processed between 2031 and September 2026 with style confidence scores between 0.74 and 0.77. The band between the old threshold and the new one. Cases that would have been assigned under the 0.74 model (above the old threshold) but were below-threshold under the 0.77 model.
The search ran for eleven seconds. It returned 341 cases.
She looked at the number. 341 cases had been processed in the boundary zone created by the 2031 threshold adjustment. UNCAT-7832 was one of them. She had no way to know, without reviewing each of the 341 records individually, how many had received the UNCAT outcome and how many had been assigned or flagged. But 341 was the population she had not known existed until today.
The October 3 test was not about UNCAT-7832 alone. It was about whether the current system handled 0.75-confidence cases as the 2031 protocol required — assigned, then audited — or whether some processing path still produced the UNCAT outcome that UNCAT-7832 had received in 2041. If the October 3 batch caught UNCAT-7832 and assigned it, the current system was behaving as designed. If it produced another UNCAT, the anomaly was structural and still present.
Either outcome told her something about the 341.
She made herself tea. The kettle was on the windowsill because she had moved it there in March to make room for the second monitor. She filled it without thinking about the second monitor. The working group's email was answered. She had what she had needed.
The face-up page in her notebook had the epistemological question written on it: what does rigorous pattern work mean inside a classification system whose foundations she did not set? She had written it Tuesday evening and decided it was a first-order node. It was still a first-order node. The threshold answer had not resolved the question. It had specified it.
She wrote underneath the question, in smaller letters: 341 cases in the boundary zone. The question is about them too.
The ISF classification system had a documented zone where the threshold boundary had moved between specification and implementation, and 341 cases had been processed in that zone, and UNCAT-7832 was one of them, and she now knew what the October 3 test was actually a test of. That was not a philosophical finding. It was a technical one. The question of what rigorous pattern work meant inside the system was still open. But she had enough now to work with.
October 3 was 16 days away.
She did not open the ISF interface again that afternoon. The tea was good. The deployment timeline question was one she would send in a follow-up request, probably next week, after she had thought about what exactly she was asking. Not today. Today the boundary zone existed, 341 was the number, and the face-up page had its second line.
She turned it face-down. Not because the question was closed. Because she had been looking at it since Tuesday and she knew what was on it.
There was a version of this research she could have done where the October 3 result was the answer: if UNCAT-7832 escapes the anchor set, the system has a structural gap; if it gets assigned, the gap is historical. That was the version she had been running in her head since August. Simple binary. Useful for a first pass.
The working group's email had made that version obsolete.
The simpler version assumed the threshold was stable. It was not. The simpler version assumed the protocol deviation, if any, was operational rather than principled. It was principled — documented, deliberate, grounded in processing capacity requirements that the 2028 specification had not anticipated. The simpler version did not account for the possibility that the 2041 UNCAT outcome was not a classification failure at all, but a processing-mode anomaly: a case processed by part of the system that had not yet completed its migration from one protocol to the other.
The more accurate version of the research was: UNCAT-7832 is one of 341 cases in the 0.74–0.77 confidence band. Its 2041 UNCAT outcome is anomalous under the post-2031 protocol that should have been governing the system at that point. The anomaly may be explained by incomplete deployment of assign-then-audit across all processing nodes. The October 3 test will show whether the current system reproduces the anomaly or has corrected it.
This was not a worse finding than the one she had started with. It was a more specific one.
She had been working with the ISF classification project for three years. When she began, her primary interest was in the style-extraction methodology: how the system determined which stylistic patterns constituted an anchor, what made an anchor stable, what the confidence score actually measured. She had written two papers on the methodology and had a third in progress. The UNCAT-7832 thread had started as a methodological question — why does this specific case fail to anchor? — and had become, over the past six weeks, a question about protocol history. Why did the case not anchor in 2041? What was the system doing in 2041?
The answer was: the system was in a state of transition. The 0.77 threshold was new as of 2031. The assign-then-audit protocol was new as of 2031. The deployment of both changes across all ISF processing nodes was not necessarily complete in 2041. UNCAT-7832 had been processed in that transitional period. The outcome it received — UNCAT, no assignment, no audit trail — was consistent with being processed by a node still running on the pre-2031 model, where a below-threshold case (0.75 < 0.77) with insufficient match scores would be UNCAT-flagged.
She wrote a new line in her notebook: ISF deployment timeline for full assign-then-audit adoption — check next week. A follow-up request to the working group or to the governance archive directly.
The question of what rigorous pattern work meant inside a system she had not designed the thresholds for had been on the face-up page since Tuesday. It was still the right question. The working group's email had given her two specific numbers (0.77, 34%) and one document reference (ISF-GOV-2031-IMPL-047) and a count (341). These were not answers to the epistemological question. They were the material with which the question could be asked precisely.
October 3 was in sixteen days.
The ISF terminal on the second desk was running its standard Thursday afternoon processing cycle: 2,200 cases through the afternoon batch, preliminary results in the daily digest at 6 PM. UNCAT-7832 was not in this batch. It was in the October 3 batch, which would process a scheduled segment of the UNCAT queue. She had put the alert in her calendar in August. She would see the result in the October 3 evening digest.
She turned the face-up page face-down.
The tea was finished. She refilled the kettle. Outside the window, the late-afternoon light was doing something particular to the building across the street — angling under the overhang in a way that only happened in mid-September. She had noticed it the year before and made a note. She noticed it again.
She went back to work.