Eval diff / 100% local
You ran your evals twice.
Now what?
Drop a baseline run and a candidate run (JSONL, JSON array, or CSV). Get a case-by-case answer: which cases broke, which recovered, which scores moved, and which are just flaky. Files are parsed in your browser and never uploaded.
3 embedded runs: baseline, candidate, and a rerun for flake detection.
1. Runs
Accepted formats
.jsonl- one JSON object per line; blank lines are fine..json- an array of objects, or an object wrapping one underresults,cases,records,rows,data,examples, oritems..csv- header row required; quoted fields and commas inside quotes are handled.- Each record needs a case id (for joining) and an outcome (boolean, pass/fail string, or numeric score). The mapping in step 2 is auto-detected and can be overridden.
- Duplicate case ids within one run: the last record wins.