Reproducibility Artifact for Zero-Knowledge Proofs on Resource-Constrained IoT: An Empirical Measurement Study of UltraHonk on ARM64 Edge Devices

This record contains the version-pinned reproducibility artifact supporting the manuscript “Zero-Knowledge Proofs on Resource-Constrained IoT: An Empirical Measurement Study of UltraHonk on ARM64 Edge Devices.” The artifact documents native ARM64 measurements of UltraHonk proof generation using Noir and Barretenberg on two generations of Raspberry Pi: a Raspberry Pi 4 with a Cortex-A72 processor and 4 GB of RAM, and a Raspberry Pi 5 with a Cortex-A76 processor and 8 GB of RAM. The package includes: • the pre-specified Raspberry Pi 4 measurement protocol;• the declared corrective Raspberry Pi 5 campaign records;• analysis-ready per-run CSV datasets;• per-prover cgroup v2 memory measurements;• thermal and performance-counter logs;• independent kernel out-of-memory witnesses;• per-run verifier logs and cryptographic digests;• the Noir circuits, compiled bytecode and verification keys;• Docker and measurement-harness definitions;• canonical Python analysis scripts and notebooks;• the exact figures embedded in the associated manuscript;• SHA-256 manifests and consistency-validation scripts. For the stress circuit on the Raspberry Pi 4, two concurrent provers completed across thirty rounds, whereas three concurrent provers triggered a global kernel out-of-memory event. The exact aggregate memory peak of the failed three-prover round was not observed. On the Raspberry Pi 5, four concurrent provers satisfy the conservative sum-of-peaks admission rule. Five-prover batches completed twenty-nine of twenty-nine analysed rounds, yielding an exact one-sided 95% lower per-round reliability bound of 90.19%. All 145 proofs produced during those five-prover rounds were accepted by the pinned Barretenberg verifier, and the independent kernel witness recorded no out-of-memory event. At six concurrent provers, none of the sixteen analysed rounds completed, and the kernel witness recorded twelve global out-of-memory kills. Across the 372 recorded Raspberry Pi 5 process rows, 326 proof artifacts were produced and accepted by the pinned verifier. The remaining 46 recorded runs produced no proof artifact because the prover had been killed. One subsequently initiated six-prover round was censored by a verifier-side structured reference string retrieval failure before any process row was emitted; this round is documented separately in the artifact’s censoring ledger. Within the controlled Pedersen-chain circuit family, proof-generation peak memory exhibited a descriptive log-log scaling exponent of approximately 0.84. Peak memory was substantially more stable across the measured conditions than wall-clock proving time, which was sensitive to thermal and system-level interference. The archive enables independent regeneration of the reported numerical analyses and figures from the retained measurement records. It does not provide a bit-for-bit reconstruction of the original container build environment, a synchronised system-wide memory trace for the five-prover operating point, or all 326 per-run proof byte streams. Verification during the campaigns was not hermetic because the measured toolchain retrieved its common reference string over the network. Measurement data, logs, figures, protocol and documentation are licensed under the Creative Commons Attribution 4.0 International license. Software components are licensed under the MIT License. The LICENSE file included in the archive defines the applicable license for each component.