Scheduler Equivalence and Performance Trade-offs under a Common LLM-Serving Contract — Complete Reproducibility Artifact — m3-v1.0.0

Complete saved-evidence artifact supporting manuscript commit 42170fdcdc7c4b68ae2639b319a4892d097532a0: 5,760 CM-0.1 cells (144 pilot reuse), 134,000 synthetic statistical studies with reuse provenance and 7,200 final endpoints, 72 budget-study cells, 18 urgency certificates, reviewed fixed-policy empty-start equivalence proof, code and analysis tools. Includes 384 TIME01 variable executions paired with reused baselines and 1,440 derived ABL01 controls (zero new simulations). Original raw provider traces and confidential reviewer materials are excluded. Synthetic timing is not hardware-calibrated. Reproduces saved-evidence analyses; no claim of turnkey fresh-provider-to-full-campaign reproduction or physical GPU validation.

Release m3-v1.0.0, promoted without scientific changes from independently verified m3-v1.0.0-rc3. The canonical TAR contains 38,186 files, 3,898,019,840 bytes; SHA-256 dafaa0df527176a9c54e0c40f5e5e4d59cebbce052c034e4df4f0f5cd5278f89. Original provider traces are excluded. Frozen preparation-time metadata inside the immutable archive records its provenance and does not describe current public availability.

Public companion records: curated source https://github.com/SoroushVahidi/llm-serving-scheduler-robustness-benchmark/releases/tag/m3-v1.0.0; authoritative result-table dataset https://huggingface.co/datasets/SoroushVahidi/llm-serving-common-contract-m3; Kaggle mirror https://www.kaggle.com/datasets/soroushvahidi/llm-serving-common-contract-m3. The table mirrors contain identical underlying result data and are not independent experiments. The separate older LSSP release is not overwritten.