Answers are grounded in this paper and cite their sources.
Emissions figures in ship weather routing rest on regulatory and market data that change with new legal acts and markets. Maintaining them by hand is slow and prone to plausible errors. This paper presents the Data Steward, an agent built on a large language model that maintains these data in EcoRouter, a workbench for ship weather routing. The agent changes a value that calculations use only through an approved proposal. Deterministic tools take over the tasks that the model performs unreliably, such as reading area boundaries from legal texts, and checks run as workflows of small steps with an outcome for every item. The contribution to artificial intelligence is the design of an agent that maintains regulatory reference data under human control, together with a method for evaluating such an agent. The engineering application is the maintenance of the regulatory data on which a ship-routing tool computes emissions and regulatory costs. In a controlled test with 40 errors injected into the data that a full check covers, the Steward corrected 36 to 38 in three runs, against 20 to 28 for the same model alone with web search. In three full checks of the live data, 96 to 99 % of judged items were correct. A person who took no part in building EcoRouter found 57 of the 59 approved values correct. The results show that a commercial language model can maintain regulatory reference data with few errors once the tasks it performs unreliably are moved into deterministic code. Preprint, not peer-reviewed. Version 1. The records of the tests are published as a dataset: 10.5281/zenodo.24251511. Use of AI: during the preparation of this work the author used Claude Opus 5.5 (Anthropic) in order to correct the text, check its grammar, to create the tables and figures from the author's records, and to check the statements about EcoRouter against its source code. The same assistant also helped to run tests 1 and 3, to judge their results against the acts and publishers, to prepare the person's sheet for test 2 and to aggregate the results of the three tests. It did only the tasks that the author gave it, under the author's control. The author reviewed and edited the content as needed and takes full responsibility for it. Language models are also part of the research itself: Section 5 of the paper states which model the Data Steward ran on and which AI assistant judged its results.
No takes yet. Share an insight, caveat, or question.
Yauheni Kisialiou (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: