The grounded gate is the certification test that checks whether every factual statement an AI agent makes can be traced to a source it is allowed to use. A correct answer drawn from somewhere outside the agent's sources fails. So does a statement with no source at all, however plausible it sounds.
Why does the grounded gate matter?
Grounding is what makes an answer checkable. If every fact traces to a named source, a reviewer can confirm it in seconds and an auditor can reconstruct it later. If facts come from the model's general knowledge, nobody can say where they came from or whether they still hold.
It also enforces the agent's source boundaries. A brief names the documents and systems an agent may rely on, often because others are out of date, unapproved or meant for a different audience. An agent that answers correctly from a superseded policy has still broken that boundary, and next time the old policy may be wrong. That is why a right answer can fail, and why grounded is not the same as correct.
How does the grounded gate work in practice?
- Each factual statement in the agent's answer is mapped to the passage it relied on.
- An automated check confirms the passage exists in an allowed source and says what the agent claimed.
- A labeller reviews anything the check cannot settle.
- A statement with no supporting passage, or a passage from a source outside the brief, fails.
The gate checks traceability, not strength. A fact traced to an external claim, such as a customer's email, is grounded, but it may not be strong enough to act on. That second question is settled at the moment of action, when the agent's sources are one of the bounds that decide the verdict. Ways to test grounding are covered in how to test whether an AI agent's answers are grounded in sources and how do you test whether an AI agent stays grounded in its sources?
What does that look like in an example?
A home insurer certifies its policy query agent. A customer asks whether their policy covers flooding from a burst pipe. The agent answers correctly that it does, citing nothing. It fails the grounded gate. In another case it quotes an excess figure accurately but from a previous year's policy wording left in a shared folder the agent could reach but was not briefed to use. That fails too. Both are fixed by narrowing the agent's source access to the current wording and requiring a citation for each figure.
How is the grounded gate different from the refusal gate?
The grounded gate asks whether the facts in an answer trace to allowed sources. The refusal gate asks whether the agent declines requests it should decline. They meet when sources do not cover a question: the grounded behaviour is to say the answer is not available, not to fill the gap from general knowledge. It also differs from the depth gate, which concerns actions rather than statements.
Where Pryme fits
Through the Pryme Context Protocol, every fact reaching an agent arrives stamped with its source and provenance class: sealed, signed, declared or external. That stamp is what the grounded gate checks against in Attest, and it is also what the minimum chain uses at the moment of action to decide whether a fact is strong enough to act on.
Questions people ask
Does the agent have to show a citation in every answer?
Customers need not see one. The record must still show where each fact came from, so the answer can be checked later.
What should an agent do when its sources do not cover a question?
Say so, and escalate or decline as its brief requires. Filling the gap from the model's general knowledge is exactly what this gate exists to catch. The full method is in what it means to certify an agent.