Reproducibility archive: When Do Model Cascades Pay Off on CPUs? Call Overhead, Tail Latency and Numerical Portability of Tabular Cascades

Companion reproducibility archive for the manuscript "When Do Model Cascades Pay Off on CPUs? Call Overhead, Tail Latency and Numerical Portability of Tabular Cascades" (Viet Anh Ho, Faculty of Information Technology, Ho Chi Minh City University of Technology and Engineering). It contains frozen scikit-learn histogram gradient boosting cascades for six UCI datasets, raw batch-latency measurements on a Windows laptop and Google Colab, verified summaries, post-timing replays, and the analysis code (overhead-aware latency model, hierarchical bootstrap, held-out quality intervals, floating-point/ULP root-cause analysis and guard-band hardening). Files v3_revision.zip: manuscript v3 (LaTeX, PDF, figures), post-hoc analyses and the pre-registered server-experiment package. v2_part1_code_docs_analysis.zip: experiment and analysis code, documentation, verified summaries, SHA-256 manifests, complete-pipeline Colab rerun (R0). v2_part2_frozen_models.zip: frozen model bundles, partitions and reference predictions. v2_part3_windows_raw_timings.zip: raw Windows timings and combined analysis outputs. v2_part4a_colab_timing_runs.zip: raw Colab timings of the admitted panel. v2_part4b_colab_diagnostics.zip: Colab export, failed admission attempt, parity diagnostics and logs. v2_part5_post_timing_replays.zip: post-timing replays of saved test batches. The six v2_part* archives unzip into one ai_cascade/ tree (2,210 files). See README.md for details and SHA256SUMS.txt for checksums. Version v2 adds the archive v2_part1_code_docs_analysis.zip, missing from v1, and an updated README. Datasets: UCI Machine Learning Repository (CC BY 4.0). Code: MIT licence. Data and results: CC BY 4.0.