Load a k-eval results file to compare how different agent configurations performed across a benchmark dataset.
Drop a .detailed.jsonl file anywhere on this page, or click the button below.
Accepts .jsonl or .json — produced by k-eval run
{{ loadErr }} Check that the file was produced by k-eval run and is not empty.
{{ loadMsg }}
These samples show the largest gap between conditions — the best place to understand why one condition outperforms another. Click any row to inspect.
| # | Question | Score gap Score gap: best condition score minus worst condition score on this sample. A larger gap means conditions disagreed significantly — good samples to investigate. | |
|---|---|---|---|
| #{{ s.idx }} | {{ s.q }} | {{ s.delta.toFixed(2) }} |
Choose a sample from the list on the left. Each sample shows the agent's response and the judge's scoring for every condition.
Tip: use ↑↓ arrow keys when the list is focusedClick a condition card to view its agent response and judge reasoning below
{{ typeof tc.tool_input === 'object' ? JSON.stringify(tc.tool_input, null, 2) : tc.tool_input }}
Statements in the agent's response that the judge could not verify against the reference answer. These may be hallucinations or out-of-scope claims — inspect the agent response above for context.