Skip to main content

v1.2.0

Releval 1.2.0 (1 October 2026) adds AI judges that grade with TypeSafe's Jev, a model that answers with a probability for each grade rather than generated text, and takes you back to the page you asked for once you sign in.

Security​

  • Security hardening of sign-in. Upgrading is recommended.

Features​

  • AI judges can use TypeSafe's Jev. Add a TypeSafe provider with its API key, then create a judge on it. A Jev judge's prompt template is the question Jev is asked about each result, with a description of what each grade means for every scale; a new judge starts from a ready-made one, and the template is checked against every scale when you save it. Jev reads text only and grades one result at a time.
  • Judgments from a Jev judge show how confident Jev was and the probability it put on each grade, where other AI judgments show their reasoning.
  • Testing a TypeSafe provider or a Jev judge reports the version of the model that answered.

Fixes​

  • Opening a Releval link while signed out, from a bookmark, a shared link or a Slack notification, takes you to that page once you sign in, instead of to the dashboard.

Upgrading​

Back up your database first. This release adds database migrations, and migrations are forward-only.

docker compose pull
docker compose up -d

Migrations are applied at startup by default. To apply them ahead of the new version instead, run the image's migrate command and then start the app with MigrateOnStartup=false, as Upgrading describes:

docker run --rm -e ACCEPT_EULA=Y \
  -e ConnectionStrings__Postgres="..." \
  releval/releval:1.2.0 migrate

migrate --script prints the SQL first if you would rather review it, and migrate --list shows which migrations a database has already applied.

If 1.1.0 keeps serving while migrate runs, its AI providers, AI judges and AI judging stop working from the moment the migration finishes until 1.2.0 replaces it. The rest of Releval is unaffected.