An agent's answer can cite a real source and still be wrong. Grounding asks whether the answer is supported by the documents the agent relied on. Correctness asks whether it is the right outcome for this customer, under the rules that apply today. Certification has to score the two separately, because passing one says little about the other.
How can a grounded answer be wrong?
Grounding is usually checked by tracing each claim back to a passage in a source. If the passage exists and says what the agent says it says, the answer is grounded. That check is necessary. It is not sufficient, and the gap between the two shows up in predictable ways:
- The source is out of date. The agent quotes last quarter's fee schedule accurately. The quotation is faithful; the fee is wrong.
- The right document, the wrong clause. A policy has a general rule and an exception. The agent cites the general rule, which is grounded, for a customer the exception covers.
- The source is right, the facts are misread. The agent correctly quotes the eligibility rule but misreads the customer's account opening date, so it applies the rule to the wrong situation.
- The source itself is weak. A claim drawn from an emailed attachment or a web page is grounded in that text, but the text has external provenance and nobody answers for it.
- Something is left out. Every sentence is supported, but the answer omits a condition that changes the outcome.
Each of these passes a grounding check. Each one gives the customer, or the person approving the draft, a wrong answer with a citation attached. The citation makes it more persuasive, not less.
Can a correct answer be ungrounded?
Yes, and it matters almost as much. A model can produce the right answer from general knowledge or a lucky guess without any support in the firm's sources. On the day of the test it scores well. But nothing ties the answer to the firm's own rules, so when those rules differ from the general case, the agent will confidently give the general answer.
An ungrounded correct answer is a coincidence. You cannot certify an agent on coincidences, because you cannot predict when they will stop.
Why must certification score them separately?
Because the four combinations call for different responses, and a single blended score hides which one you have.
| Correct | Wrong | |
|---|---|---|
| Grounded | The answer you want | Source, version or application problem; fix the sources or the brief |
| Ungrounded | Right by coincidence; will fail when your rules differ | Hallucination or reasoning failure; the most visible error |
When you certify an agent, the grounded gate checks support for each claim, and correctness is checked against the outcome an expert labelled for that case. They are separate measurements with separate floors. An agent that is well grounded but often wrong usually has a source or versioning problem. One that is often right but poorly grounded is relying on the model rather than on you, and its success will not survive a change in your policies.
The practical side of the grounding check is covered in how to test whether an AI agent's answers are grounded in sources, and the question of whether grounding holds up over time in how you test whether an AI agent stays grounded in its sources.
Who decides what "correct" means?
Grounding can be checked mechanically to a degree: does this passage exist, does it support this sentence. Correctness cannot. It needs a person who knows the process to decide, before the run, what the right outcome for each case was. That label is the reference the agent's answer is scored against.
This is why correctness is the more expensive half of certification and the half most often skipped. Teams that only check grounding are, in effect, letting the agent's sources stand in for an expert. If the sources are complete, current and unambiguous, that works. Most firms' sources are not all three.
Where does provenance come in?
Provenance narrows the gap between grounded and correct without closing it. A claim grounded in a sealed source, such as the policy library or the core ledger, is far more likely to be correct than one grounded in an external document, because the sealed source is maintained by the system of record and is tamper-evident. Knowing the class of each source lets you weight what grounding is worth.
But even a sealed source can be applied wrongly. The policy library may be perfectly current and the agent may still cite the wrong clause for this customer. Provenance tells you how much to trust the source; it does not tell you whether the agent used it properly.
This is one reason testing a vendor's agent on the vendor's material does not settle the matter. As the piece on whether certifying a vendor is the same as certifying an agent argues, grounding and correctness are only meaningful against your own sources and your own labelled outcomes.
What should you change?
Look at your current evaluation and ask whether it would catch an answer that quotes the right document and reaches the wrong outcome. If the only check is "did the agent cite a source, and does the source say that", it will not. Add expert-labelled outcomes for each case, score grounding and correctness separately, and report the four combinations rather than a single figure.
Treat a grounded-but-wrong result as seriously as an ungrounded one. It is the failure customers are least able to spot, because it comes with a reference.
Where Pryme fits
In Pryme, Attest scores grounding as one of seven gates and checks each answer against the outcome the customer's own experts labelled. Through the Pryme Context Protocol, every fact the agent relies on carries its provenance class, so a reviewer can see whether an answer rested on a sealed policy or an external document.
Questions people ask
Is a high grounding score a good sign on its own?
It is a good sign that the agent uses its sources. It says nothing about whether those sources were current or applied to the right facts. Read it alongside correctness, never instead of it.
Can grounding be checked automatically?
Partly. Matching claims to passages can be automated to a useful degree, with sampling by people. Whether the right passage was chosen for the customer's situation still needs expert review.
What usually causes grounded-but-wrong answers?
Out-of-date or duplicated documents are the most common cause, followed by exceptions the agent did not apply. Versioning the source library fixes much of the first; edge cases in the certification set expose the second.