Replication package for "Deployment Parity of Object Detectors on Edge CPUs: An Empirical Study Across Four Inference Runtimes and Two ARM Microarchitectures"

Replication package for a measurement study of what changes when the same trained object detector is carried across inference runtimes and ARM microarchitectures on edge CPUs. Contains every derived measurement from a campaign of 184 parity cells and 193 timing cells across four devices spanning two ARM microarchitectures (Cortex-A72, Cortex-A76), four inference runtimes (ONNX Runtime, OpenVINO, MNN, ncnn) and six trained networks: layered parity results, per-image output digests, fault-injection sensitivity and the detection floor, version and operator bisection of a quantising compiler regression, the half-precision intervention, latency confidence intervals, and the gate outputs. Also contains the full pipeline, self-contained reproducers for the compiler regression, the generator that produces the paper's tables and figures, and a harness that re-derives every number printed in the paper from the shipped data. The 500 COCO val2017 evaluation images are not redistributed; the package carries the sampling seed, the image ids and a SHA-256 digest for every image, together with a script that rebuilds the identical subset from the official release and verifies each file against those digests. Board IP addresses and key paths are replaced by placeholders; host names are kept as published in the paper. Code is MIT licensed; data is CC BY 4.0.