Select a case to inspect its output and assertion evidence.
Structured output review · browser-based
Catch regressions before your next AI release.
Compare a baseline and candidate run, apply a few explicit assertions, inspect case-level differences, and download a review packet. Your run data stays in this browser session.
No account, network save, autosave, or model API. Example records are synthetic.
Import two runs
Paste or choose one JSON or JSONL export for each run. Import applies both together.
JSON uses an array of records. JSONL uses one record per line. Each record needs a stable id and an output; include input and baseline expected when available.
See format examples and download templates
JSON
[{"id":"case-01","input":{"question":"..."},"output":{"answer":"..."},"expected":{"answer":"..."}}]Download JSON templateJSONL
{"id":"case-01","input":{"question":"..."},"output":{"answer":"..."},"expected":{"answer":"..."}}Download JSONL templateTemplates are synthetic starting points. Records are limited to 2,000 rows per run.
Choose assertions
Point to an output field. No code or prompt editing needed.
/answer/items/0/nameUse JSON Pointer paths; ~1 means “/” and ~0 means “~”.“Equals expected” reads its value from the baseline record’s expected field. Assertions run against candidate outputs. Editing runs, labels, or assertions clears previous results and reviewer decisions.
Review differences
Load the example or import runs to start a comparison.
Keep a portable copy
Save a project JSON file and reopen it in a later browser session. Nothing is stored automatically.
Example data is loaded. Changes are not saved automatically.
Possible future edition · unavailable
Would a shared team workspace help?
A proposed team workspace could keep shared review history and CI connections. It is not available yet; the free browser review and local CLI remain usable without an account.