The Annotation

He opened the case at 9:04am.

He had not planned to. He had planned to clear the two standard cases first — both straightforward, both above 90%, both sitting at the top of the queue since Monday — and then look at the wetland case again at eleven, as he had looked at it on each of the previous six Wednesdays, and then defer it, as he had deferred it on each of those Wednesdays, into the afternoon and then into the evening and then off his screen.

He opened it at 9:04am instead and looked at it without reading it for a moment and then began to read.

The confidence interval was 84.7%. He had known this for twenty-two days. He read the ecological projection — net habitat gain, core system, 2045–2060 — and then the paragraph he always stopped at: Western buffer zone: current viability assessment 31%. Projected trajectory without intervention: continued saline intrusion, estimated loss of productive habitat 60–70% by 2060. Recommended action: managed surrender. Rationale: resource reallocation to higher-viability zones produces greater ecosystem return per unit of management effort.

Managed surrender. He read the phrase again. He had read it eleven times. It still meant what it meant.

He read the Mawali community section. It was three pages. He had read it eleven times. Seven villages, largest settlement Banshtala, population approximately 4,200. The livelihood breakdown: fishing, some rice cultivation on the higher ground, and the honey harvest — the Apis dorsata colonies worked in March and April by the Mawali's traditional harvesters, the mawalis properly speaking, who read the movement of the colonies through the surviving canopy and had done so for longer than there had been a system to survey the delta. Relocation timeline: 24 months. Disruption level: moderate. Livelihood transition support: eligible for Authority resettlement programme BR-7, employment retraining, community infrastructure grant.

He closed the case file.

He sat for a while.

He opened a new document. He looked at the blank page. He had written the annotation in his head eleven times — had drafted sentences in the shower, on the train, lying awake at eleven at night — and none of those sentences were in the document because it was blank.

He began to write.

He wrote: This is a note for the record. It is not a counterargument to the recommendation. I am going to sign this case. What follows is an account of what I understand and what I don't, because I think those two things are different and the record should say so.

He stopped. He was not sure the record should say so. He left it in.

He wrote about the confidence interval. 84.7% — high enough, under the Authority's guidelines, to sign without annotation, or with only a brief notation of any specific concerns. He wrote: I am annotating this anyway because the 15.3% is not distributed evenly across the decision. It concentrates in the parts of this case that the model has less data on. I know which parts those are. I want to say which parts those are.

He listed them. The tiger movement data — sparse, the core-to-buffer transit routes tracked partially, two sensors offline in the southern corridor since October and not yet replaced. The honey harvest productivity projections — derived from regional analogues, not from Banshtala-specific observation, because the system does not have the granularity. The relocation outcome data — based on comparable resettlements, three of which he had reviewed in detail, two of which had resolved reasonably well over a decade and one of which had not, and the one that had not was not in this projection because the outcome period used was eight years and the deterioration in the third case had become visible at year nine.

He wrote that down. The eight-year outcome window. The year nine data that was not in the model.

It was 10:17am.

He thought about the morning in Banshtala.

It had been six months ago — a site visit at the beginning of the Priority 1 designation, standard protocol, two hours on the ground before the boat back. He had walked the embankment edge and then into the buffer zone's outer margin, a younger officer from the regional coordination team beside him, and the officer had pointed out the landmarks in the case documentation — the village cluster, the canopy cover percentage, the sensor stations reporting to the central network.

There had been a woman working at the edge of the tree line. She was preparing smoke equipment — a smoker, the kind used to calm the bees before the harvest, the bellows checked methodically, the fuel tamped. She was not preparing for the harvest then; the harvest was two months away. She was maintaining the equipment. She was reading the canopy above her in the same motion — looking up while her hands worked, the way people look at something they are responsible for, not at something they are observing.

He had watched her for perhaps thirty seconds. The officer had been speaking about the sensor station coverage. He had said yes at an appropriate moment.

He wrote about her now. He did not know her name. He wrote: On the site visit in August I observed a member of the Mawali community maintaining harvesting equipment at the edge of the buffer zone's outer tree line. She was also reading the canopy. These were not two separate activities. The system's productivity projections are based on regional analogues. They are not based on what she knows. I cannot put what she knows into this annotation because I do not have it — I watched her for thirty seconds and I do not know what she was reading in the canopy or what it told her. I am noting this absence specifically because the model's 84.7% confidence does not account for it and I think the record should say so.

He stopped and read what he had written. He did not know her name, and the record would not either, and he understood that this was not a failure of the record but a condition of it.

He wrote: You do not have a variable for what leaves with the people. I am adding this to the record. It is not a counterargument to your recommendation. It is an acknowledgement of what the recommendation cannot see.

It was 12:44pm. He ate at his desk — he had stopped going to the kitchen at noon when the other Adjudicators were there because he had found himself unable to join the conversation, not from unfriendliness but from the distance it took to hold where he was in the annotation.

He wrote about the model's recommendation. He wrote that he thought it was probably right. He wrote: The core system is worth saving. The buffer zone's 31% viability is not a permanent floor — it is a trajectory, and the trajectory goes one direction without the freshwater diversion resources that are currently split between the buffer zone and the core. I understand this. The model is likely correct that concentrating management effort in the core produces greater ecosystem return. He wrote: I have read the three comparable resettlements. Two of them resolved reasonably well. I think this one probably will also. I am prepared to be wrong about that.

He wrote about what he was prepared to be wrong about.

This section took a long time. He wrote it carefully because it was the section he would read again, in six months or in three years, when the outcomes became available, and he wanted the record to show that he had understood the weight of it before he signed. He was not looking for absolution. The annotation was not absolution — it was accountability, which was different, which was the thing the system had been routing to him for two years without his knowing that was what it was doing or why.

He did not know that that was what it was doing.

He wrote.

At 1:58pm he wrote the section on what he could not know. This was the hardest section. He wrote: I cannot know what the buffer zone's ecosystem will look like in 2060. I cannot know whether the third comparable resettlement will resolve well or whether it will not. I cannot know what the woman with the smoker knows about the colony movement in that canopy. I cannot know whether the resource reallocation to the core system will produce the projected ecosystem gains or whether there will be a variable the model has not accounted for — a different silting pattern, a salinity event, a failure mode the system has not yet classified. He paused. He wrote: I am signing on the basis of what I do know: the trajectory of the buffer zone, the viability of the core, the model's track record on comparable decisions, and my own assessment of what the annotation above contains that the model's confidence interval does not. He wrote: I could be wrong. That is what my signature means. It means I am the person who is accountable for being wrong, if this turns out to have been wrong.

He read this back.

He left it in.

He saved the document at 3:11pm.

He did not sign the case. He had known, working through the annotation, that he would not sign it today — that the annotation was the thing he had to do first, and that the signature was tomorrow, and that he could do tomorrow's thing tomorrow. He closed the case file and saved it. He closed the annotation and saved it.

The system processed the save. A small notification appeared at the top of the screen, below the document title, in the same pale banner as the reversal notice three weeks ago.

A1 classification confirmed.

He looked at it. He had seen the label before — in the reversed decision's superseded record, the inaccessible metadata field, the documentation search that had returned access level required. He looked at it for a moment. The banner was the same pale colour as every other banner. The font was standard. The label was the same one he had not been able to find a definition for.

He was very tired.

He closed the terminal and put on his coat and picked up his case and walked to the lift. His reflection moved toward him in the darkened terminal screen and then the screen went to sleep and it was gone.