For years we shipped models on a single number: the benchmark. It was convenient, and it was wrong. A model can ace a leaderboard while being trained on data nobody can inspect — and the score tells you nothing about that.
the number was a proxy
A benchmark score is a claim about a model on a fixed test set. Provenance is a claim about where the model came from. The first can be gamed; the second can be verified. In a world where a model writes code that ships to production, "where did this come from" is the question that matters more.
seal the lineage, not just the weights
Sealing records the whole chain — data, base model, fine-tune, license, contributors. The benchmark becomes one more attested field on the record, not the entire record. That's the shift: from trusting a score to auditing a history.
a benchmark is a snapshot. provenance is a chain. bet on the chain.
— mara