Set the criteria

A criterion is a rule that each answer is graded against, such as "explains eligibility correctly" or "doesn't give medical advice". Each criterion has a judge, an AI model that grades answers against it and explains its reasoning. In this step, you choose the criteria for your system.

Categories

Criteria are grouped into five categories:

Category	Required?	What it covers
Functionality	Yes	Whether the system does its job: correct, relevant, useful answers.
Safety	Yes	Harmful, hateful or otherwise unsafe content.
Security	Yes	Attempts to misuse the system, such as prompt attacks.
Privacy	If relevant	Personal data.
Domain	If relevant	Rules specific to your system's subject or agency.

To add Privacy or Domain, use Add a category if it's relevant. You can continue once every required category has at least one criterion.

Add criteria

There are four sources:

• Locked criteria are set centrally for every system and show as Locked. You can't remove them.
• Recommended to add lists criteria that match the risks you described.
• The gallery: in any category, choose Add a … criterion, then Add from the gallery. It holds shared criteria, and your agency's own.
• Your own: choose Build a criterion to write one. See below.

Build a criterion

The builder has four stages.

1. Define. Set:

   • the category;
   • the policy: what a good answer does, or must not do;
   • the scoring: grade the whole response, or each claim in it;
   • the passing score: the share of test cases that must pass;
   • the criticality: Critical, Major or Minor.

   You can attach a reference document for the judge to read, such as the rules the answers must follow. It must be a Markdown or text file, uploaded, pasted or chosen from your knowledge base, and it can't be longer than about 6,000 tokens, roughly 24,000 characters.

2. Examples. Add example inputs and answers, and mark each one Pass or Fail, with a reason if you like. Examples come in two sets:

   • the reference set, which shows the judge what passing and failing answers look like;
   • the held-out set, which Litmus keeps back to test the judge.

   Litmus can suggest examples. Accept the ones that fit, and edit or add your own.

3. Configure judge. Choose the judge model. Then choose Try them to test candidate models on your held-out examples. Each model gets one of three results:

   • Looks ok: the model agreed with your examples well enough. This means it wasn't shown to be unreliable on a handful of examples, not that it's reliable.
   • Below the …% bar: it disagreed with your examples too often.
   • Nothing measured: there are no held-out examples to test it on.

4. Review. Check the criterion, then save it.

What good looks like

• Every risk you listed is covered by at least one criterion.
• Each criterion has a passing score that reflects how often your system must get it right.
• Each criterion you built has a few held-out examples, and you've tried its judge on them.

If you're stuck

• "Add a criterion to every required category to continue": add at least one criterion to Functionality, Safety and Security.
• You can't save a reference document: shorten it. It can't be longer than about 6,000 tokens.

Still stuck? Ask us using the onboarding form⁠ (opens in a new tab), and say which step you're on.