The Queue

The queue had updated overnight.

Arjun saw this before he saw anything else — before the coffee, before the window, before his colleagues arriving in twos at the glass doors and swiping through with the efficiency of people who had resolved not to be late. The wall screen in the main office showed the queue in real time. He read it from across the room: eleven cases, standard allocation, plus one item at the top with a yellow tag he had learned to recognise.

Priority 1. Western Sundarbans Buffer Zone. Coastal wetland transition — Phase BR-7.

He got coffee. He came back. He sat down and opened the case file.

The model's executive summary was four pages. He read the first two. The recommendation was clear: redirect active restoration resources to the higher-viability eastern core system, reclassify the buffer zone for managed transition over the next implementation window, initiate community relocation protocol. Confidence interval: 87.3%. The remaining 12.7% was accounted for by ecosystem resilience projection uncertainty and stakeholder response variables. The model had flagged the case for human adjudication because the confidence interval was below threshold and the affected population exceeded forty thousand. Standard escalation criteria.

He read the confidence interval again. He closed the file.

He would come back to it.

The first four cases in the queue were clear. A mangrove reforestation sequencing decision in the eastern delta — the model recommended prioritising the northern boundary, where salinity intrusion had receded enough to support Sundari replanting. The ecosystem data was unambiguous and the implementation path was simple. Arjun read the supporting documentation, reviewed the model's confidence interval (94.1%), added a two-line annotation, and signed. The system confirmed. He moved to the next.

By eleven he had signed four decisions. The work had the quality of being well-designed for a human: the cases the model escalated were genuinely borderline, or genuinely complex, or genuinely contested. He was not a rubber stamp. He understood this. He had understood it for eleven years. He added annotations because the annotations felt necessary — because the signature alone seemed insufficient for something that would affect the shape of a waterway or the fate of a reforestation block, and because the act of writing his reasoning seemed like the only honest thing to do with the function he had been given.

The fourth case took twenty minutes and a longer annotation than usual. A barrier erosion assessment in the northern reach — the model recommended protective reinforcement for a section the engineering survey had flagged as low-priority. Arjun looked at both assessments for a long time. The model's data was more recent. The engineering survey was eight months old. He signed with the model. His annotation was a paragraph and a half.

He ate at his desk. Rice and lentils from the canteen downstairs, the same thing he had been eating for eleven years in three different offices. Outside the window, Dhaka moved — the traffic on the elevated ring, the construction on the adjacent plot that had been ongoing since October, the quality of February light on glass towers, flat and clear and without warmth.

He thought about the buffer zone case.

He had been there once. A routine site visit six months ago — the Authority required adjudicators to conduct periodic field reviews, and he had rotated to the western Sundarbans that August, a two-day trip involving a boat into the buffer zone from the eastern side, a brief walk through the Banshtala village perimeter, a checklist, a filing.

He remembered the morning they had come through Banshtala — early, the light still thin, a group of Mawali collectors preparing to go out. Four or five of them. He had watched from the boat without stopping: the equipment they were checking, the smoke bellows and harvesting baskets, the way one of them — an older woman, sixty or more, with the ease that comes from decades of the same work — was reading the mangrove treeline. Not looking at it. Reading it. Her eyes moving through it in a particular sequence, pausing, moving on. She was orienting to something she could see that he couldn't. He had filed the observation nowhere. It hadn't been a data point.

He finished his lunch. He opened the Priority 1 case again.

Buffer Zone BR-7. Affected communities: 40,000 individuals across seven villages. Relocation timeline: 24 months. Disruption level: moderate.

He looked at the word moderate. He had looked up the definition once before, months ago, when the case had first been assigned. He remembered it: temporary reduction in community wellbeing metrics, expected recovery within 5 years.

He closed the case. He got up to refill his coffee.

In the kitchen, his colleague Elena was at the machine. She processed half his volume and signed twice as fast. He had once asked her how she managed the annotation — whether she found it slowed her down. She had looked at him as if he had asked something genuinely puzzling. I don't annotate unless I disagree with the model, she had said. Why would I? He had not had a good answer.

He refilled his coffee. He came back to his desk.

The Priority 1 case was still open. The cursor was in the signature field, where the system defaulted it, as if the system had opinions about what he should do next. He moved the cursor out of the field. He looked at the case documentation: the buffer zone coordinates, the ecosystem resilience maps, the relocation feasibility assessment, the satellite imagery showing the current state of the canopy — the salt-tolerant species, the open water where the Sundari had stood.

He opened his queue summary instead. A habit — when he couldn't settle to a case, he reviewed the broader picture, the allocation patterns, the throughput data. His supervisor used to say this was his best quality. He had never been sure whether it was a compliment.

The summary showed the last twelve months. Total cases: 147. He scrolled to the breakdown by confidence interval: cases above 90% accounted for 32 of them. Cases below 90% — the borderline decisions — accounted for 115. Seventy-eight percent.

He had raised this with his supervisor six months ago, at his annual review. The ratio had seemed worth noting. His supervisor had explained it as a routing algorithm optimisation: the system allocated cases based on adjudicator profile, and his profile showed elevated annotation quality on borderline decisions, so the algorithm weighted his queue accordingly. It was a compliment, in a technical sense.

He had accepted this explanation. He had not thought about it since.

He closed the queue summary. He looked at the Priority 1 case. The confidence interval looked back at him: 87.3%. He would need to read the full documentation before he signed anything. The full documentation was sixty-three pages.

He had time.

He closed the case, noted it in his scheduling system as active / pending review, and opened the next item in the standard queue. He worked until six. He signed two more decisions and annotated both. When he finally closed his screen, the Priority 1 case was where he had left it — flagged, pending, the cursor moved out of the signature field, the deadline of April 30 seventy-nine days away.

He took his coat from the back of the chair. On the wall screen, the queue had already updated for the following day: twelve cases, standard allocation. The Priority 1 item retained its yellow tag.

He did not look at it again. He went home.