A Case-Derived Five-Stage Decision-Audit Proposal for Feature-Evaluation Decisions in Analytics Systems: One Card-Transaction Illustration

Abstract

Supplementary Code S1 contains selected non-sensitive analysis runners, contracts, manifests, gate definitions, figure source, and aggregate outputs. The EXP-67/68 runners, contracts, and aggregate outputs are included but were not replayed or re-executed during this package build. For EXP-68, stage1_q.csv records only the observed-Q threshold; the refit-Q 97.5th percentile was not computed, so the full Section 3.4 gate was not evaluated. Verification is limited to the EXP-63 and EXP-66 aggregate replay (five CSV comparisons) and four PNGs within the figure check's stated mechanical scope; this verified subset does not define the entire code selection. Full model reproduction is incomplete. The package excludes original transaction data, original public-dataset files, customer-level predictions, person identifiers, and identifier-bearing metadata. Aggregate outputs derived from public datasets remain subject to the applicable dataset terms of use. MIT applies only to original package code and documentation, not to third-party material or datasets. A bounded static inspection of 20 external-data adapters found zero copied code; this finding is limited to that static check and is not a general provenance audit. This research was supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF), funded by the Ministry of Education (RS-2021-NR060140), and by the ANCHOR program through the Seoul ANCHOR Center, funded by the Ministry of Education (MOE) and the Seoul Metropolitan Government, Republic of Korea (2026-ANCHOR-01-020-04). The version 3 static selection adds the EXP-94 through EXP-100 runner and aggregate categories, analysis-plan YAML and checksums, plan-rebinding receipts, feature and claim registers, the shared revision core and path module, and the claim-value checker. EXP-99 is represented by aggregate cohort flow. These added artifacts were statically screened but were not scientifically re-executed during packaging. Version 4 adds the EXP-75 and EXP-77 runner code, their selected summary-level outputs, the missing local import package file, and the Historical Routing-Rule Record document. The focal refit runner and its selected aggregate outputs were already included and remain included. The three claim-register source snapshots are refreshed. The claim-value checker writes the current claim-evidence map with the same header and identical file bytes. The separate claim-map builder is excluded because it requires a review module outside the public scope. The added runners were statically screened but were not scientifically executed during packaging. No restricted inputs, customer-level outputs, checkpoints, or additional draw-level files are added. Upload and publication records are handled separately.