Leaderboard screenshots without model versions are souvenirs. Teams that pin model IDs, prompt hashes, and dataset revisions can explain a score drop. Teams that do not will argue about ghosts.

Model version, prompt hash, dataset revision, score delta — Eval hygiene: pins make deltas meaningful.

Minimum pin set

Exact model string / snapshot ID, not a marketing family name.
Prompt template hash and tool schema version.
Dataset revision and a short note on what changed since last run.

Public habit: refuse to compare unticketed scores in procurement meetings.

Sources: Public LLM eval methodology notes; MLOps experiment-tracking habits; buyer RFP reproducibility asks. Soft-lander essay — no product claim.