[Re] Independent Python replication of E. coli starvation-stress reporter analyses Partial computational replication and sensitivity of source-defined analyses

We independently replicate the exported numerical analyses underlying Figure 6B/C/D of Gervais et al.’s Escherichia coli starvation and oxidative-challenge study and examine how the observational hierarchy affects their interpretation. The recovery panel contains 1,153 software-defined cycles, including 245 unresolved outcomes; 1,145 recovery IDs appear in the pre-challenge reporter table, with 1,139 carrying usable measurements of both reporters. Treatment arms form two disjoint sets of experiment dates, making a four-arm comparison distinct from a within-date contrast. Under unrestricted binary completion, the pooled CAM-after-minus-CAM-throughout contrast remains positive by at least 0.210 percentage points, but both September 2022 dose strata admit an ordering reversal; the June 2023 dose strata remain positive. Equal weighting of 37 source-identifier partitions is a separate descriptive estimand: its small cross-date order reversal requires 130 simultaneous, oppositely assigned unresolved labels. We reproduce the author’s exported reporter summaries, recover panel-specific sample definitions, and provide runnable source checks, exact finite-record sensitivity intervals and figure reconstruction. The work clarifies when a pooled software-level ranking is robust, and what assumptions are required to compare it across observation sets and source dates.