BOA Constrictor: Benchmarking Neural Backbones for HEP Data Compression

The LHC at CERN generates 30 petabytes of physics data every year. Current compression tools like LZMA treat this data like any other file and miss the statistical patterns unique to particle physics. The BOA Constrictor solves this by training a neural network to predict the next byte more accurately inside an entropy coder. The original BOA paper proved the concept works (4.03x compression on CMS data vs LZMA's 3.22x), but two gaps remain: Mamba doesn't install on Python 3.12, and nobody has compared alternative architectures under fair conditions. This project will build a benchmark suite comparing GRU, LSTM, MinGRU, Transformer, and Mamba under parameter-matched conditions. It will then optimize the best backbone using INT8 quantisation, pruning, and distillation, producing Pareto frontier plots showing compression ratio vs model size. Finally, it will explore physics-informed priors as an auxiliary loss to improve predictions. All code and results will be contributed back to the upstream BOA repository.