Reproducibility checklist

Complete this checklist before promoting a result as independently reproducible.

Dataset and protocol

• Record the exact Amazon Reviews source, category, version, and checksum.
• Publish all filtering, ID remapping, split, and sampled-negative parameters.
• Preserve the evaluated user/item counts and exclusions for every run.
• State clearly whether evaluation is sampled-negative or full-catalog.

Training

• Record the Git commit SHA, immutable config, Python and dependency versions.
• Record hardware, runtime, seed, batch size, stopping epoch, and checkpoint hash.
• Run at least three independent seeds for each learned model.

Reporting

• Publish per-seed metrics, mean, standard deviation, and confidence intervals.
• Save raw prediction/rank files sufficient to recompute every table.
• Generate README tables from checked-in result files with a script.
• Separate synthetic smoke-test output from real-data performance results.

Suggested release layout:

results/<release>/
├── manifest.json
├── environment.txt
├── dataset_checksums.txt
├── predictions/
├── runs/
└── tables/