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

| File | Contents ||---|---|| `v3_revision.zip` | Manuscript v3 (LaTeX + PDF + figures); post-hoc analyses: overhead-aware latency model (`analysis/costmodel.py`), hierarchical bootstrap (`analysis/bootstrap_ci.py`), held-out quality intervals (`quality/`), log1p/ULP root-cause and tie-exposure/guard-band hardening (`ulp/`); pre-registered server experiment package (`server_experiment/`, `PLAN_E3.md`). || `v2_experiment_archive.zip` | Frozen-artifact experiment (V2): experiment and analysis code (`q2_revision/code/`), frozen model bundles and partitions (`q2_revision/prepared/`), raw Windows timings (`q2_revision/runs_clean/`), verified Colab exports incl. the failed admission attempt (`q2_revision/colab_results_extracted/`), diagnostics, post-timing replays, verified summaries (`q2_revision/analysis_*`), complete-pipeline Colab rerun R0 (`replication_colab/`), dataset provenance (`data_sources.md`) and SHA-256 manifests. |