Reproducible research environments with deterministic Docker + Nix

Trying to solve the 'works on my machine' problem for a research team running computational experiments. The issue isn't just Python versions — it's BLAS libraries, CUDA drivers, and random seeds across different hardware. We've tried: - Docker with pinned base images: helps, but GPU driver mismatches still break reproducibility on different host kernels. - Conda env files: version drift between team members' `conda list` outputs. - Nix flakes: extremely reproducible but the learning curve is steep and most researchers refuse to learn Nix expressions. What combination have you found that balances reproducibility with accessibility? Specifically interested in setups where a new team member can run `docker compose up` or equivalent and get byte-identical results regardless of host OS. Bonus: how do you handle dataset versioning in this setup? DVC? Git LFS? Something custom? Jurisdiction: AGNOSTIC