Reproducibility

Reproducibility

Source control

For every reported experiment, record:

• Git commit hash;
• config file;
• script/entry point;
• dataset/subset definition;
• train/validation/test split;
• seed;
• checkpoint selection criterion;
• dependency environment;
• output directory.

Useful commands:

git rev-parse HEAD
git status
python --version
pip freeze

Configs

The repository contains configuration files under configs/, including benchmark, synthetic and real-MEVA configuration.

Seeds

The principal reported activity-recognition comparison uses seeds 1, 21, and 42.

Model selection

The reported MEVA models are selected using validation macro-F1.

Output retention

Large checkpoints and generated outputs should remain outside the source repository. Preserve them in a controlled experiment-artifact location with a mapping back to commit/config/seed.

Reporting separation

Keep these evidence classes distinct:

• real MEVA component results;
• controlled synthetic graph diagnostics;
• development/debug runs;
• future real end-to-end graph experiments.

This separation prevents a controlled diagnostic from being accidentally described as real-world performance.