Evidence explorer
What changed in the decision?
Compare recorded conditions with candidates aligned by meaning. Trace each value back to the frozen experiment.
LogitTrail was formerly named Mini Jev. The original records and download identifiers retain that name.
Start with an observation
Browse the retained items
Both sides always use the same source item, task, and model. Original task text is not included in this artifact; item IDs and source hashes identify it. This is not a model ranking.
Candidate comparison
Probability within the candidate set
A and B share a 0–100% scale. Δ is B − A in percentage points. Values are aligned by semantic key, even when presentation order changes.
Concentration measures how peaked a candidate distribution is; it is not the probability of being correct. Score expectations assume equally spaced stages. Recorded examples do not establish human usability or a causal model-size effect.