A meter query can run without an error and still return the wrong answer. It might read the right meter for the wrong hour. That distinction matters if an operator is about to investigate an alarm.

In the OpenMeter preprint, Carlos Martín-Abellán and Mark Kröll test a system that turns a question into a database query and then explains the result. One component plans the work, another writes and retries the query, and a third interprets the readings. The authors use a sandbox with read-only execution and test on roughly 7.5 million rows from an industrial partner. The local experiments ran on premises; the separate cloud experiments used data with identifying information anonymized before transmission.

On a 9-billion-parameter model running locally, the full pipeline lifted the relaxed execution score from 36.3% to 42.4% compared with the best single-agent setup on that measure. That measure allows extra columns as long as the reference values are present and the row count matches. On the stricter test, which also checks column count and order, the full pipeline reached 32.2%. These are averages from five runs of 49 questions. On the larger cloud model, the plain single-agent baseline had the best relaxed score. The local gain makes those extra steps worth testing on smaller models. The scores still leave plenty for an operator to check.

For a pilot, give the system a small set of questions with results someone has already checked. Include a missing interval or two meters with similar names. Have the operator see the time range, units and rows that produced the answer. The useful result is a reading the operator can check before acting on it.