Finding
All 51 available state summaries reconcile in each of 2023 and 2024. Twenty-six summaries in 2025 exceed the 0.11-minute reconciliation tolerance. The workbench uses 2024 and retains the newer discrepancies for investigation. A separate predictive candidate also fails its approval checks.
Why it matters
Check a release before treating it as an upgrade. Analysts can inspect comparable utility groups using the reconciled year while pursuing the unresolved source differences.
Who could use this
An analyst reviewing utility peers and deciding which source releases can inform due diligence.
Data and scope
Six official EIA reliability workbooks from 2020–2025: 11,552 method-specific records across 1,010 utility-state keys, with original workbook, sheet and row lineage.
How the analysis works
Python ETL, SQLite aggregation, method-consistent peer cohorts, state-total reconciliation and report-year-separated model testing. Utility-cluster bootstrap uncertainty checks whether an apparent accuracy gain is convincing.
Limits of the evidence
The predictive candidate also fails promotion: its improvement interval crosses zero and nominal interval coverage falls short. That diagnostic uses the quarantined source year. No engineering recommendation or actual upgrade saving is established.