Which Code Where? Code, MobileNetV2 checkpoint and per-seed results for exact per-layer ECC allocation in quantized DNNs

Code, model checkpoint and per-seed results for the study Which Code Where? Exact Per-Layer ECC Allocation for Quantized DNNs Under a Check-Bit Budget. Each weight layer of an INT-8 network chooses one protection option (none, parity, SECDED (72,64), DEC-TED (79,64), optionally with 2- or 4-way bit interleaving) under a single check-bit budget. The choice is solved as a multiple-choice knapsack problem over a two-context value table measured by fault injection, followed by an evaluation-in-the-loop refinement, and compared at equal cost on held-out images and fault draws with uniform codes, a SECDED ranking rule, greedy search, a genetic algorithm, the cross-entropy method, QMIX and VDN. Contents: eccmarl_code.tar.gz - source code (study driver: eccmarl/codealloc_study.py; manuscript, analysis and rendering scripts: paper2/). codealloc_results.tar.gz - per-seed results.json files (ResNet-18/CIFAR-10 and MobileNetV2/CIFAR-100; burst shares 0-0.6 at BER 3e-3 and 0.3 at 1e-3; five seeds; budgets 0.25-2x SECDED) with value tables, allocations, held-out per-draw accuracies and pod logs. mobilenetv2_cifar100_s42* - the MobileNetV2/CIFAR-100 FP32 checkpoint (seed 42), its INT-8 export and training metadata. The ResNet-18/CIFAR-10 checkpoint is the public model jaeunglee/resnet18-cifar10-unlearning on Hugging Face. Decoder sensitivity (v1.1). codealloc_results.tar.gz also holds 36 runs that repeat the three main fault settings (iid, burst share 0.30 and 0.60 at BER 3e-3) for both networks with seeds 0-2 under (i) a real (72,64) Hsiao syndrome decoder for SECDED words and (ii) reloading, instead of zeroing, words with detected uncorrectable errors (fault-model names *_hsiao and *_reload).