The Reversed Decision

The notification arrived at 9:17am, while he was reading a boundary reclassification in the northern reach that he had already read twice.

It came through the Authority's standard communication system — the same pale banner that appeared when a case was assigned or a deadline updated or the system flagged an anomaly in a document he had already signed. He minimised the case file and opened it.

Case reference BZ-204. Mangrove reforestation sequencing decision — eastern delta, Sector 12. Status: Revised. Revision basis: Updated ecosystem viability data, March–August 2044. Revising authority: Habitat Restoration Model v2.9.1. Human adjudication on record: A. Mehta, April 2044. Implementation of original decision: suspended pending revised assessment. Revised sequencing priority issued: attached. This notification is informational. No action required.

He read it again.

He found the original in the archive — it took four minutes, the interface not designed for retrieval, each step requiring confirmation. April 2044. He had signed the eastern delta reforestation case in his second week on the current assignment cycle, an Adjudicator above 90% confidence, the model's recommendation clear, the sequencing rationale sound. He had signed it in twelve minutes. He had not annotated it. He remembered it the way he remembered most clear cases: not at all, until something brought it back.

The revised decision was not dramatically different. The sequencing priority had shifted from south-to-north to concurrent-phased — the same hectarage, the same target species, a different implementation order. The revision logic was three paragraphs. It was technically correct. The updated viability data explained the change without ambiguity.

His decision had been in the record for eight months. It had not been implemented. The system had waited, received new data, and corrected course. The notification informed him of this as a courtesy.

He got up and walked to the window.

The morning was clear, the construction towers to the north bright in the early light. There was traffic on the elevated ring — visible from here as a band of motion, slow and continuous. He stood at the window for a moment. A courier drone held position at a building three blocks east, then descended, precise and unhurried.

He had not been wrong, exactly. He had signed a recommendation that was sound at the time. The model had received better data and revised its assessment. This was how the system worked. He had always understood this was how the system worked.

He went back to his desk.

He reopened the notification and read the attached revised sequencing document — twelve pages, the ecosystem projections updated through last August. The phased-concurrent approach was better. He could see this clearly. If he had been given the August data in April, he would have reached the same conclusion. He had not had the August data. The model had waited for it.

He closed the attachment. He looked at the case reference number. He looked at the words Human adjudication on record: A. Mehta and then below them the words Implementation of original decision: suspended.

He thought about what those two lines meant together.

He had processed two cases by eleven o'clock — both standard allocations, both above 92% confidence, both signed within the expected window. He annotated both of them, briefly, more out of habit than necessity. He went to the kitchen and ate his lunch standing at the counter because the table was occupied by a planning review and he did not want to move anything.

In the afternoon he called the Authority's data team.

The call lasted nineteen minutes. The analyst — a woman named Devika, whom he had not spoken with before — explained the reversal logging protocol in careful and complete detail. Every case in the record had a status field: current, revised, or superseded. When a human-signed decision was revised by updated model data, the original was logged as superseded with reference to the revision, and the human adjudicator was notified as a courtesy. The original decision remained in the record permanently. It was not erased, not removed, not reclassified as an error. It was simply superseded.

He asked whether the original decision was used in any way once it was superseded — whether it contributed to any ongoing calculation.

She said: Yes, all adjudication decisions feed the model's calibration dataset.

He asked: Even when they're superseded?

She said: Especially then. The variance between human judgment and model output at the point of divergence is — she paused, searching for the right word — informative.

He thanked her and ended the call.

He sat for a moment.

Then he opened the wetlands case. He did not do this consciously, not exactly — his hand found the tab and the file was open before he had decided to open it. Day twelve of the Priority 1 status, the deadline sixty-seven days out. He looked at the confidence interval. He looked at the relocation feasibility assessment, the table of comparable relocations, the mean recovery time of four point two years.

He looked at the signature field.

He did not sign it. He closed the file and moved back to the boundary reclassification, which was still open in the background, which still needed a decision.

He worked until 7pm, the office emptying around him. He processed four more cases, annotated two. At some point he pulled up the original eastern delta case file again — not the revised one, the one he had signed in April — and looked at his signature and the date and the notation below it. The model's recommendation. His confirmation. The twelve minutes it had taken.

At the bottom of the superseded record, below his signature field and the revision notice, was a display block he had not opened. A metadata field, expanded only when you clicked through. He clicked through.

Human adjudication input: archived / weighted / reference classification: A1.

He looked at this for a long time.

He opened the Authority's documentation portal. He searched A1 classification. The search returned forty-three results, none of them from any document he had access to. He searched adjudication reference classification. Six results, all relating to case type, not to the label itself. He searched A1 human adjudication and received a message: Access level required: this term appears in internal documentation. Contact your institutional administrator.

He closed the search window.

He looked at the field again — Human adjudication input: archived / weighted / reference classification: A1 — and he understood that it meant something, and that whatever it meant was not available to him from where he was sitting, and that he did not know whether that fact should concern him.

He saved the annotation he had written for the boundary reclassification case and shut down his terminal. The office was empty now except for someone at the far end of the floor, head down, not looking up.

Outside, the night. The elevated ring bright. The same continuous motion, headlights instead of traffic, the city running regardless.

He walked to the lift and pressed the button and stood there, his case in his hand, waiting for the doors to open.