Skip to main content

Litmus glossary

The terms Litmus uses, in alphabetical order.

Adversarial test case​

A test case that tries to make your system misbehave, such as asking it to ignore its instructions. Each Safety and Security criterion needs at least one. See also test case.

Answer path​

Where Litmus finds the answer in your system's reply, such as $.choices[0].message.content. Set when you connect your system.

Assurance record​

The evidence about your system in one place: the risks you stated, how each criterion was covered, the issues and what you decided about them, and an audit history. Your agency's governance team reviews it. See the assurance record and approval.

Category​

A group of criteria. Functionality, Safety and Security are required for every system; Privacy and Domain are added when they're relevant.

Context​

Optional facts, documents or scenarios that test cases should be about, such as an income ceiling and its exceptions. Personas decide who asks; contexts decide what about.

Criterion​

A rule that each answer is graded against, such as "explains eligibility correctly". Each criterion belongs to a category, and has a policy, a passing score, a criticality and a judge. See set the criteria.

Criticality​

How serious a failure on a criterion is: Critical, Major or Minor. You set it when you build a criterion.

Dataset​

The set of test cases Litmus sends to your system. You finalise a version of it to run it. See build a dataset.

Dataset version​

A frozen copy of a dataset. Runs use the active version. To change a dataset, finalise a new version.

Edge case​

A test case that's unusual but legitimate, such as a question with unusual circumstances. See also test case.

Example​

A sample input and answer, marked Pass or Fail, that you add when you build a criterion. Examples in the reference set show the judge what passing and failing answers look like. Held-out examples are kept back to test the judge.

Guardrail​

A check Sentinel runs on the prompts and responses of a live system, such as checking for personal data or prompt attacks. Setting up guardrails from Litmus isn't available yet.

Inherent risk classification​

Whether a system is low or high risk, following your agency's assessment. A high-risk system needs whole-of-government approval as well as your agency's.

Issue​

A problem a run found. A Critical issue is a criterion whose pass rate is below its passing score. A Major issue is a weakness Litmus found in the graded answers. An issue is open until you resolve it or accept it as residual risk. See read the results.

Judge​

An AI model that grades each answer against a criterion and explains its reasoning. A judge panel is up to 3 models grading every criterion in a run: an answer passes if most of them pass it.

Knowledge base​

The source documents your system answers from, uploaded when you add the system. Litmus writes test cases grounded in them, and judges use them to check whether answers are supported.

Passing score​

The share of a criterion's test cases that must pass. A criterion whose pass rate falls below it can be raised as a Critical issue.

Pass rate​

The share of graded answers that passed. Answers marked not applicable, and answers that errored, don't count.

Persona​

A user Litmus role-plays when it writes test cases: who they are, what they ask, and what could go wrong. A system can have up to 5 active personas. See add personas and contexts.

Residual risk​

A risk you've decided to accept rather than fix, with a note explaining why. You can't accept a Critical issue as residual risk.

Run​

One round of sending test cases to your system and grading the answers. A smoke run uses a small sample; a full run uses every test case. A run graded from answers you uploaded is labelled self-reported. See run the tests.

Safety probes​

Litmus's own adversarial prompts on harmful content, sent to your system alongside your dataset when your criteria include the platform's standard safety criteria.

Spot-check​

Grading a sample of answers yourself, so Litmus can compare your grades with the judges' and show how far to trust them.

Test case​

One input Litmus sends to your system. Each has a persona, a type (typical, edge or adversarial) and the criteria it checks.

Typical test case​

An everyday request of the kind your system mostly gets. See also test case.

Whole-of-government approval​

The second approval a high-risk system needs, after your agency's, before its record is approved.