Skip to content

What the judge is

A judgebot answers Magic: The Gathering rules questions asked in plain language. It gives a short ruling backed by quotes from the Comprehensive Rules, Scryfall rulings or the card’s current Oracle text. Each quote is checked against its source before the answer is shown, so an answer cannot cite a rule number the model invented. Ask from Discord with /judge, from a web app, or from the command line. Use brackets like [[Full Card Name]] when a nickname could mean several things.

Every judgebot is its own Discord application, run by the community that uses it. Each has its own spend cap, judge role and bot token, so the person who chose the model pays for the questions. The design decisions page has the reasoning.

Setting one up takes a compose file, a model API key, and a Discord application you create in the developer portal:

  1. Requirements and first run: the database, the data loads, and the web app on your own machine.
  2. Create the Discord app: the portal steps, with links into Discord’s documentation for each.
  3. Production deployment, when it should stay up without your laptop.

You can see the judge work before you have a bot token: Try it without Discord brings up the web app and the command line. Once your bot is in a server, /help explains the commands, and the Discord commands page has the details.

For every question answered, an instance stores:

  • the question text and the answer
  • the channel or web session it was asked in, and whether it came from Discord, the web app or an agent
  • on Discord, the asker’s user id and username
  • whether it was asked private
  • the ids of the rules, rulings and cards it was answered from.

A Discord question asked with /judge-private is stored too, but only the operator can see it: it is never shown to anyone else, never read as another question’s history and never used as an example. A failed call is kept with what the model sent, until a later failure finds it older than 30 days, so the operator can diagnose it. When someone presses a rating button, the instance stores their Discord user id, the score and, for Partially correct or Incorrect, the reason they give in its form. A rating given in the web app stores the browser tab’s session id instead of a user id, the name the rater typed if any, and the address and browser it came from, which are cleared after FEEDBACK_IP_DAYS (default 90). It stores nothing else. The bot receives only its own slash commands and button presses, never channel messages.

Every answered question not asked private is listed on the web app’s public feed, with the Discord asker’s username, unless the operator turns it off (JUDGE_FEED=false). Servers and channels are never shown there.

/forget deletes a user’s ratings and reasons and removes their id and username from every question they asked. A private question loses its text too, and so does a failed one. A public question keeps its text, which the channel saw. A question from the web app or an agent has no user id.

Sample answers shows what comes back, and Evaluation how it scored on a hard question set.

Every citation is checked against its source before it is shown, and every rule number named in the answer’s text must also be cited. That rules out invented rule numbers and misquoted text, but not a wrong conclusion drawn from correct quotes. Verify anything that matters at a tournament with a human judge. The bot declines tournament-policy questions (the Magic Tournament Rules and Infraction Procedure Guide).