Auditing Historical Features at the Cold-Start Boundary: Public-safe Replication Package for Predictive Performance and Decision List Instability in Public Procurement Risk Scoring

Anonymity notice for peer review: This record has been uploaded under an anonymous author identity to preserve double-blind review. Author names, affiliations, and ORCID identifiers have been intentionally omitted from all package files, DOI metadata, and creator fields. The corresponding published version, if accepted, will replace this anonymous record with a fully attributed release. Abstract Public managers using algorithmic risk scores allocate scarce administrative attention, not merely predictions. We test whether actor and buyer-supplier histories change which cases enter a review queue while delivering stable cold-start and proper-score gains. Using 707,725 Colombian SECOP II contracts in a prospectively locked 2022 outer test, we compare a contract/process/context baseline with a full model adding actor and relationship histories under a prospectively pre-specified joint practical-and-inferential rule. None of four confirmatory history-reliance tests was supported. Ranking contrasts were sometimes positive, including after common-support weighting, but proper-score evidence was absent, negative, or scale-sensitive, and elastic-net and LightGBM replications disagreed in direction. Outcome-blind history substitutions moved mean ranks by 4.65-8.00%, changed 2.15-7.02% of the top-5% review list, and worsened Brier skill in every scenario. Historical features therefore redistributed administrative attention without demonstrated stable predictive benefit. We propose cold-start composition, decision-list stability, and proper-score quality as auditable performance attributes. Package contents: This public-safe replication package (33 files, 817,576 bytes) contains code, tables, figures, and documentation sufficient to reproduce all confirmatory results and robustness checks reported in the manuscript. See 00_README and 12_SHA256_MANIFEST.txt inside the archive for the file inventory and integrity hashes. Keywords: public management; public procurement; artificial intelligence; cold-start evaluation; decision-list stability License: Documentation and tables are released under CC BY 4.0. Source code inside the archive is released under the MIT License (see LICENSE files inside the ZIP).