docs/SCALING_PROTOCOL.md

HOI Scaling Protocol v1

Primary axis

Dataset × model tier × seed. Primary dataset: CIFAR-10. Extension: CIFAR-100. Seeds: 7, 17, 27.

Controlled variables

Use the checked-in config, 32×32 input for official CIFAR runs, AdamW, no unreported augmentation, and the benchmark's test evaluation. If a change is scientifically useful, give it a new config name rather than overwriting an official tier.

Required metrics

Accuracy, NLL, train wall seconds, samples/second, peak allocated GPU memory, parameter count, device name, PyTorch version, AMP status, train/test sample counts, epochs, seed.

Scaling extensions

Recommended high-compute experiments: width {128,256,512}; depth {6,8,12}; modes {8,12,16}; spectral rank {16,32,64,full}; resolution {32,64,128} on an appropriate dataset. Change one axis at a time before factorial sweeps.

Fair comparisons

Add CNN/ViT/FNO baselines only with the same data and evaluation protocol. For efficiency claims, provide both parameter-matched and measured-throughput comparisons. Report failed/OOM runs because the feasible frontier is itself a scaling result.