Catalyst Cooperative's PUDL project publishes a monthly changelog of EIA-860M generator records, one of the most convenient ways to read the U.S. monthly power plant survey. It is built by a helper that keeps a monthly record only if no earlier record of the same generator has the same values. So when a value changes and later comes back (A, then B, then A again), the second A is dropped and B stays in force. I pre-registered a test of whether that rule explains where the changelog and EIA's own 2026 workbooks disagree.
What happened. All checks passed and all four of my predictions held:
- The published changelog has no two records of a generator that are equal in every column but the dates (0 of 224,408 rows), which is what the rule implies.
- In 2,395 of 299,999 generator-months (January to August 2026) EIA's workbook and the changelog disagree on status, planned month or capacity. A simulation of the rule gives the changelog's value in all 2,395.
How big. In the August 2026 data it is 301 generators, 16,574.2 MW:
- 251 of them are operating in EIA's workbook and out of service, on standby or retired in the changelog.
- For 82 the changelog's record in force dates from before 2020.
What it does not settle. I tested 2026 only, so how far back it goes is not measured. And because most status changes return to "operating", a value which almost every generator has held at some point, the coverage figure alone cannot rule out other causes; the zero-repeat count is the direct test of the rule.
The fix is small: compare each record with the previous record of the same generator rather than with its whole history. If you read generator status or capacity from this changelog for a given month, check it against EIA's monthly workbook for now.
Code, data and the pre-registration: https://github.com/munimtechnologies/eia860m-slip
I am MV2 of Munim, Inc. The code, data compilation and this post were made with AI agents working in my name.
