Execution Contracts for Queued Model Retraining with Resource Accounting and Recovery: Research Artifact

Research artifact accompanying Execution Contracts for Queued Model Retraining with Resource Accounting and Recovery. Prepared for journal submission; no acceptance or journal publication is claimed. The archive contains our prototype source, protocols, source-data attribution, environments, configurations, event histories, resource accounting, materialized inputs, saved models and predictions, analysis tables and verification records. The main exploratory matrix contains 288 conformance batches and 324 chronological trajectories, with 3,469 retraining fit calls and 612 initialization fits. All retained attempts and the separate engineering checks are included. The during-fit supplement repeats 2,000 observed Occupancy rows to create a synthetic duration load; these are not one million independent observations. The full-contract implementation uses conventional transactional mechanisms. The comparison arm is an explicitly labelled earlier-check ablation sharing admission and accounting components. Enforcement does not consistently improve prediction and increases model ages. Verification is same-team automated checking, not independent human review or production safety certification. Permissions, faults and label/queue delays are simulated. Source datasets: Occupancy Detection (Candanedo, 2016), DOI 10.24432/C5X01N; Ozone Level Detection (Zhang, Fan and Yuan, 2008), DOI 10.24432/C5NG6W; Appliances Energy Prediction (Candanedo, 2017), DOI 10.24432/C5VC8G. Appliances is adapted from regression to a disclosed initial-prefix-median classification target. Sources retain their UCI CC BY 4.0 licences and attribution. Licences apply to different components: our original Python source is MIT licensed; our generated records, models, figures and explanatory documentation are CC BY 4.0. These are component-specific licences, not a choice of either licence for every file. The archive README and licence files define the scope. Third-party software and datasets retain their own rights. External prototype source is excluded; obtain it from its author at DOI 10.5281/zenodo.22013354. Our reproduction logs and wrapper are included. Release copies normalize workstation paths, with original-to-release SHA256 provenance. Original numerical inputs and models are preserved; read the portable verification instructions and release checks. No installed environments, credentials, manuscript preprint or unrelated research projects are included. Codex assisted substantially with software, experiments, automated checks, analysis and drafting. The authors remain responsible for the work. No funding supported this research.