SECAI Core · phase 5 of 15
Grounding check
A verification step that tests whether the AI model's outputs are correctly supported by specified sources or evidence (such as documents, databases, or the input data itself).
The Explain card
- Plain English
- A grounding check is a verification step that tests whether a model's output is actually supported by the specified sources or evidence, such as documents, databases or the input data itself.
- Example
- After a RAG system summarises a vendor contract, a second pass compares each sentence of the summary against the contract text and flags "the vendor guarantees 99.99% uptime" because no such clause exists in the retrieved pages.
- Why it matters
- Giving a model sources does not guarantee it used them. A grounding check catches hallucinations and quiet drift before outputs reach a decision maker, and provides an auditable record of what was verified.
- Hook
- Grounding check: the model cited a source, now go see if the source agrees.
Where it sits in the deck
Phase 5: Interacting with Models: Prompting, APIs, and Retrieval
With a working model and clean data, learn the interaction layer — how users and systems communicate instructions and retrieve grounded answers.
