What Is Groundedness?
Groundedness is the property of an answer whose every statement is supported by the material supplied alongside the request. It is a relation between two texts — the answer, and the sources placed in the context it was produced from — and not a judgement about the world. An answer can be perfectly grounded in a source that is out of date, and a statement that is true but appears in no supplied passage is ungrounded.
How a groundedness score is actually produced
The measurement is taken statement by statement rather than over the answer as a whole, because support is rarely uniform across an answer. A paragraph whose first three sentences come from a retrieved passage and whose fourth came from nowhere is the ordinary case, and a single verdict over the paragraph loses exactly the sentence that was worth finding. So the answer is first decomposed into short assertions that stand on their own — pronouns resolved, qualifying clauses attached to what they qualify, one thing asserted per assertion. Each is then judged against the supplied passages and marked supported, unsupported or contradicted, and those marks are aggregated into whatever the evaluation reports.
Two details decide whether the resulting figure means anything. The first is that decomposition is where most of the error is generated: an assertion extracted with an unresolved reference, or with a qualifier left behind in the sentence it came from, will be judged wrongly however good the judge is, and the mistake presents as a judging mistake. The second is that support is not textual overlap. An assertion can be supported by a passage sharing almost none of its wording, and can fail to be supported by a passage it copies from, if the copy fuses two sentences into a claim that neither made. Judgements are therefore made by something that reasons about entailment — a trained inference model, or a separate generating pass given the assertion and the passages — rather than by matching strings.
Support that requires more than one passage is the case implementations quietly disagree about. An assertion assembled correctly from two supplied passages is grounded in the supplied material taken together and ungrounded in either passage taken alone, so an evaluation that judges each assertion against its single nearest passage will mark it unsupported and an evaluation that judges against the whole context will not. Neither is wrong; they are answering different questions. It is worth knowing which one a score came from before setting that score beside another.
A second axis of variation is what the judgement is attached to. Some evaluations ask only whether the supplied material supports the assertion, wherever in that material the support happens to sit; others ask whether the particular passage the answer cited supports it, which is a stricter question and a different one. An answer that reaches a correct conclusion from the fourth supplied passage while citing the second is grounded under the first reading and fails under the second. The strict reading is the one worth paying for wherever a reader is expected to follow a citation, and it costs more, because it requires the answer to have carried the attribution correctly as well as the content.
| Property | The question it asks |
|---|---|
| Groundedness | Is every statement in the answer supported by the material supplied with the request? |
| Factual correctness | Is the statement true of the world, whatever the supplied material happens to say? |
| Answer relevance | Does the answer address the question that was actually asked? |
| Context relevance | Was the material supplied pertinent to that question in the first place? |
| Citation accuracy | Does each citation point at the passage that supports the statement it is attached to? |
| Completeness | Did the answer omit something the supplied material contained and the question required? |
Only the first row is groundedness, and the last row is the one it structurally cannot see. An answer that drops the qualifying sentence from its source remains grounded in everything it did say, which is why an evaluation resting on groundedness alone quietly rewards an answer for saying less.
What a groundedness score does not tell you
It does not tell you the answer is true. Groundedness inherits the accuracy of whatever was retrieved: a store serving a superseded policy document produces answers that score well and mislead, and the score cannot separate that case from the good one, because it never looks outside the context it was given.
It does not tell you the retrieval worked. The score is taken against whatever happened to be supplied, so a pipeline that retrieved almost nothing relevant and answered cautiously by restating the little it got can score above one that retrieved well and reached further. This is why retrieval quality is measured separately and read first: under the same defect the two numbers move in opposite directions, and the reassuring one is the groundedness score.
And it does not survive being used as a target. A figure produced by a judge is a signal, and a pipeline tuned until the signal is high learns the shape the judge rewards — hedged sentences, near-quotation, and claims narrow enough that the supporting passage is unmistakable. That output is more grounded and less useful, and nothing in the measurement registers the second half. Where the number gates a release, the thing worth keeping beside it is a sample of answers that a person reads.
One further property is easy to overlook and awkward once it has been seen: the judgement is generally produced by the same kind of system whose output is being judged. A generating pass asked whether an assertion follows from a passage is performing reading comprehension, and it can fail in exactly the direction it has been deployed to detect — finding support in a passage that does not carry it, or missing support expressed in unfamiliar terms. That does not make the measurement worthless, and treating it as a measurement rather than as a verdict is most of the discipline here. The figure is a reading from an instrument with error of its own, which is why an evaluation reporting one is expected to say how its judge was checked, against what, and by whom.
How the Registry classifies a groundedness check
Checking each statement in a generated answer against the passages supplied with it and marking the unsupported ones is filed as data processing in this registry's classification. Both inputs exist before the task begins — an answer, and the material it was meant to rest on — and what the task produces is a judgement derived from setting one against the other.
The Registry's canonical brief for this entry is filed as:
Check each statement in a generated answer against the passages supplied with it and mark the unsupported ones.
Submitted for assessment it is
classified as Data processing, and its wording is hashed once — to
b16164ae2042a298…, the first sixteen of sixty-four hexadecimal
characters — with the wording itself never stored. The hash is what the derivation
reads.
That class's own page is /tasks/analysis.
Classification is one of three inputs. The other two are the configuration submitted with the task, and the permanent chart derived from that configuration — fixed by the model name, the training cutoff and the temperature alone, and never reading the task at all. The same ascendant, ruling planet and harmony therefore appear on every assessment a given configuration receives, whatever it was asked to do. The derivation is published in full at /method.
What this page does not claim about groundedness
Which decomposition rule splits the answer, what is allowed to count as support, and who or what makes the judgement are the three choices that decide every groundedness figure, and the canonical brief above names none of them. This registry checks no answers, holds no evaluation set, and has never scored the groundedness of anything.
The Registry does not run this task, does not inspect any system's output for it, and validates no assessment it issues against what afterwards happens. What it does is compute — from a published method, for one submitted task and one submitted configuration — a verdict and a recommended execution window. It computes neither on this page.
Questions about groundedness
- What is groundedness?
- Groundedness is the extent to which every statement in an answer is supported by the source material supplied with the request. It is a relation between the answer and that material, so it says nothing about whether the material itself is correct.
- How is groundedness measured?
- The answer is broken into short self-contained assertions, each assertion is judged against the supplied passages as supported, unsupported or contradicted, and the marks are aggregated. The judgement is usually made by a trained entailment model or by a separate generating pass given both the assertion and the passages.
- Is a grounded answer a correct answer?
- No. Grounded means supported by what was supplied. Where the supplied passage is wrong, superseded, or misapplied to the question, a fully grounded answer repeats the error faithfully. Correctness is a separate property, measured against the world rather than against the context.
- What is the difference between groundedness and hallucination detection?
- Groundedness is the property being measured; hallucination detection is the activity of finding statements that lack support. Where sources are supplied the two largely coincide. They separate where nothing was supplied to compare against, since a groundedness score then has no referent and detection must fall back on other signals.
- Can an answer be fully grounded and still be a bad answer?
- Easily. Grounding says nothing about whether the answer addressed the question, whether the retrieved material was the right material, or whether something essential was left out. An answer that restates one supplied sentence and stops is fully grounded and frequently useless.
Order an assessment for data processing tasks
The Registry issues a permanent, numbered task risk assessment for one submitted task and one submitted configuration. EUR 1.90 Standard, EUR 4.90 Extended, EUR 14.90 Full Chart, which adds the permanent chart. Assessments from EUR 1.90; machine-readable at /pricing.json.