Persistent disturbance and false extremes in the Carrington baseline

The apparent severity of a historical magnetic disturbance depends on which values are transcribed, which observing times survive and how incomplete records are compared. We reconstruct the August–September 1859 baseline and quantify these effects around the Carrington event. A discrepancy between a published Melbourne data file and its source photograph changes the August 30 daily-range rank from fourth to fifty-first among 61 dates when the source reading is substituted. Conversely, September 5 retains the largest recorded adjacent-hour declination change in the two-month Melbourne series, 27.19 arcmin, despite ranking only third by daily range. Retaining only even-hour samples reduces its recorded range by 49.2% and moves its range rank to thirteenth. Across four Russian observatories, September 5 ranks first by a normalized four-hour range when analysis is restricted to 55 complete dates; restoring usable windows from incomplete dates places it fourth or fifth among all 61 dates. An amplitude-sensitive common-decrease score also identifies a stronger value on September 24. These results distinguish sustained disturbance from a unique event signature and show how transcription, sampling and completeness policies alter the inferred baseline. The underlying source-traced archive preserves native values, derived coordinates, exclusions and correction histories. The contribution is a quantitative reconstruction and sensitivity analysis of known historical observations, including the previously documented September 5 activity. Version 1.0 is a research preprint and has not been peer reviewed. The deposit includes the paper and a source-and-code companion containing editable manuscript text, figure assets, document-building code, analysis scripts, input specifications and verification records. Forty-one retained network/reconciliation outputs and six additional comparison outputs were reproduced byte for byte on 28 October 2026. The companion supports rebuilding the document but is not a self-contained numerical reproduction package: the wider data archive and curated input tables are not included, pending source-specific redistribution decisions. Independent human transcription review remains outstanding. AI tools assisted source reading, programming, analysis and manuscript preparation; responsibility for the final manuscript remains with the author. The paper and original figure assets are licensed under CC BY 4.0. Project code is licensed under MIT. Upstream observations retain their original source terms; this deposit does not relicense or redistribute the full underlying datasets.