The database separates what was observed from how strongly the public evidence supports it. A striking demonstration can be worth tracking without being treated as equivalent to a peer-reviewed and independently replicated benchmark.
Peer reviewed
A primary research paper has passed journal peer review. The label says nothing about replication, sample size or whether the chosen benchmark is commercially meaningful; those remain separate questions.
Preprint
A manuscript is public and can be inspected, but formal peer review is incomplete. Preprints stay labelled even when the result looks plausible.
Company demonstration
A developer has published a product, demo or performance claim. The entry records what can be inspected publicly while keeping manufacturer-reported metrics distinct from independent validation.
Platform versus computation
Some entries matter because they make experiments reproducible or accessible rather than because they beat a benchmark. The database includes them when the platform materially changes what researchers can do.
Updates
Records are dated and stronger sources replace weaker ones. Raw system records are available as CSV and JSON, making changes easy to inspect or reuse.