Reproducibility corpus for the E1 experiments in "Serialization-Invariant Content Addressing for ML Model Artifacts" (Paper A).

Reproducibility corpus for the E1 experiments in "Serialization-Invariant Content Addressing for ML Model Artifacts" (Paper A). The experiments test whether a canonical structural address (κ-address) assigns the same identity to a machine-learning model across serialization-level changes — operator reordering, initializer reordering, and other transforms that alter the bytes of an ONNX file without changing the model it represents. Contents. Experiment scripts (transforms, address sweep), result tables including the headline e1_table.csv, two figures, the reproducibility protocol/runbook, and MODELS.md — a manifest of the six stock ONNX models used, each with its source URL and verified SHA-256. Reference, not redistribution. The base models (≈565 MB) are standard public artifacts from the ONNX Model Zoo and are not included; MODELS.md records the exact variant, source URL, and SHA-256 for each, and the scripts regenerate the transformed-variant corpus deterministically. This keeps the archive small (44 KB) while remaining fully reproducible: fetch the six models per the manifest, verify their hashes, and rerun. Version sensitivity. κ-addresses depend on the canonicalization ruleset version. Results here were produced with the pinned dependency versions listed in the README; a different uor_addr canonicalization version may yield different addresses. Cite the concept DOI for the latest version. Code is released under MIT; data, results, and figures under CC-BY-4.0.