Skip to content

Grounding and Citations

2 min read

Grounding means the AI bases what it writes on your Canon documents, and each piece of output carries a citation: a source tag showing which document (and section) it came from.


You don’t need to switch anything on. Every AI Test Case Generator run uses Canon unless you turn it off.

The Ground in Canon toggle sits with the other inputs in the AI Test Case Generator. Turn it off to generate from the attached requirement alone.

Turn it off when Canon is incomplete, out of date, or still processing. A half-ready Canon can be worse than none, because some output is based on your specs and the rest is not.

Your choice is remembered for you, per project, until you change it. A project with no Canon documents needs no change: the AI simply finds nothing to use.

When a document is relevant, the AI reads the whole document, not just a few matching lines.


FeatureWhat you see
Generated test cases and gapsEach draft shows one source tag per document it used. Hover over or focus a tag to see the file name and the passages (short sections) the AI read. Gaps show them too, so you know which document needs the missing detail. See Write and review test cases.
Release readinessThe AI summary cites the exit criteria it checked the release against. Without Canon, it falls back on general AI knowledge, such as “80% pass rate looks healthy”. See Release readiness.
AtlasAreas suggested from Canon show the documents they came from, so you can check before accepting them. See Propose from Canon.
OracleAnswers about your specs list the Canon documents used, with the page or section for each. See Answers from documents.

A source tag points to a place in the document: a heading (§ 4.2 Expiry handling), a page (p. 12), a table (Table 3) or a spreadsheet sheet (Sheet: Q3).


  1. Read the test case.

  2. Look at its source tag. Is this the document you’d expect to describe this behaviour?

  3. Hover over the tag and read the passage. Does it actually say what the test case checks?

  4. If it does, accept the draft. If it doesn’t, either the document is wrong or the AI went too far. Fix the document or edit the draft.


Grounding can only use what your documents actually say. If your specs describe features without rules or outcomes, the test cases will be vague too. See Canon best practices.