What supports this picture?

Source and identity

Qu, Jingbo (2026), Dataset for “A Multi-Agent AI Framework for Explainable Battery System Maintenance”, Zenodo, CC BY 4.0. Attribution and transformations identifies what this presentation changes; no endorsement by the data creators is implied. The record used here is OperationCondition_202502_All.mat → OperationCondition_202502_All → Operation35, locally indexed as #260. Its start label is 2025-02-13 11:30:00; the container lasts about 161.37 minutes. R1–R9 are recorded rack positions, not verified installation coordinates or replacement histories.

The presentation inherits a previously checked extraction. It does not reopen the original MAT or independently establish sensor validity. The source locator preserves the publisher filename, groups and inherited original-file MD5. The exact extracted I/V archive is included separately from the display packet. It is not the original raw MAT.

What was retained

64,865 current rows and 12,001 voltage rows labelled DISPLAYED. Three pre-existing segmentation-event current rows—one each for R1, R6 and R9—remain excluded from plots and calculations, but their status is preserved in the extracted archive. No temperature, SOC, cell voltage, control log or sealed record is part of this case.

Minutes are derived from the record’s start timestamp. Native points retain their own source timestamps and row indices. Unsupported time is not reconstructed, treated as rest or filled with zero. Current is measured realised current, not requested forcing; negative values indicate charge.

How paired records were formed

At the union of source timestamps, take each member’s latest row at or before that time. Keep a node only if every participating source is no older than the chosen screen: 5 s for the main reading, 0 and 10 s for sensitivity. Resolve tied timestamps using the highest source index; suppress unchanged source-index tuples.

The direct pair is built from R1 and R4 alone. Subtract their recorded values at each retained node. The nine-member reference C is built separately and is the median of nine recorded values. Neither node type is a measurement of a simultaneous physical state; equal recorded timestamps do not prove simultaneous acquisition.

Five-minute bins are half-open intervals from the container start. Their summaries weight retained nodes, not elapsed time or independent measurements. The numerical table keeps the earlier frozen bins; the display does not refit their boundaries. Current medians agreeing under 5/10 s screens is a sensitivity check, not independent physical evidence.

Selection, missingness and rendering

#260 was the next chronological record before it was opened. R1/R4 and the 80–120 minute close-up were selected after inspecting all nine members and the full record. The case is exploratory, not a blind test or evidence of recurrence.

The whole-record current overview contains every retained native current row. The detailed current panel also uses native rows, with the other seven members faintly visible. Difference panels use the existing paired nodes. All panels draw points only: no decimation, connecting lines, smoothing, interpolation or extrapolation. The close-up selects time, and its vertical scale contains all selected values. The full-record button restores complete time and full value ranges.

The cursor can inspect only a nearby eligible node inside the view, within 5 s of its selected position, independently for current and voltage. This radius is a navigation convention, not a new measurement or significance criterion. It does not carry values across a gap. Source time, age and row index appear in its readout. The close-up is [80,120) minutes; the full record includes its final retained point.

Voltage is substantially sparser. Under the main screen, 80–120 min retains 114 voltage pairs, none in 85–90 min. Their negative difference does not establish an unchanging voltage relation between nodes.

Checks and what they do not establish

The inherited reading compared event-sweep and binary-search node construction. This presentation separately checks all displayed source witnesses, reconstructs nodes from the extracted source rows, and reproduces the tabulated signs and counts. These are implementation and arithmetic checks, not an independent scientific review. Hashes establish byte identity, not physical calibration, synchrony or cause.

No empirical null for recurrence, mechanism or control benefit is claimed. No such result is emitted. The only case result is the bounded ordering of these retained recorded differences.

Run python reproduce.py beside the two .json.gz downloads and claim-values.json. Python standard library only. This checks extraction-to-display arithmetic, not the upstream MAT extraction or physical experiment.

Corrections and reuse

This edition retains the descriptive reading of the earlier local draft. A bounded public-reader check found no material wording objection on that draft; it was not a scientific validation of the underlying measurements. Version and source notes remain available for corrections and reuse.

To flag a reading error, retain the record number, channel, rack, source row, time and edition. A successor should identify the correction while preserving this edition’s evidence and limits. Keep source attribution, selection and missingness with reuse; an observation case is not a battery-control recommendation.