Grounding Validation

What it is

Deciding whether a produced answer is actually supported by a defined set of evidence supplied for that request, and whether it may therefore be delivered. It presupposes that the evidence set exists and is identified — the passages placed in the context, the rows a tool returned, the document the user attached — and it yields a verdict with a threshold and a configured consequence rather than anything a reader sees.

Supported, cited, and true are three questions

Citation produces a pointer a reader can follow, and its quality is about whether the pointer resolves, whether it is at the right granularity, and whether it survives the corpus being re-segmented; it is a feature of the answer that a person consumes. Support is a yes-or-no question asked on behalf of the system rather than the reader: given these premises, does this sentence follow, and is the proportion of the answer that follows high enough to ship. Truth is a third question entirely, answered against the world rather than against the supplied material — an answer can be perfectly supported by a document that is wrong, and the validator is correct to pass it. Catching fabrication where no evidence set was ever defined is yet another check, with different signals and no notion of a threshold over a source.

Why it matters

Every retrieval-grounded product rests on an assumption that the model confines itself to what it was given, and models do not. The result is the output such a product cannot survive: an answer in the system's own voice, formatted exactly like every correct answer, asserting something the sources do not say. A validation gate is what converts "we supplied the right documents" from a hope into a measured property of each individual response, and it is the only honest basis for telling a user that an answer could not be supported rather than quietly presenting one that was not.

Key decisions

• What constitutes the evidence set for a request, and whether the conversation history, the system instructions, and the model's own parametric knowledge are inside it or outside it.
• The unit of judgement: the whole answer, each sentence, or each atomic claim. Atomic claims are the most defensible and require a decomposition step that is itself a source of error.
• What proportion of unsupported content fails an answer, and which claim types — figures, dates, names, negations, quantifiers — fail on a single instance regardless of proportion.
• Whether the validator comes from a different model family than the generator, since one that shares its representations will tend to agree with whatever it produced.
• The consequence of a failure: suppress, regenerate with the unsupported spans named, deliver with a caveat, or escalate. Each of those is a materially different product.
• Whether validation runs synchronously on every response, on a sample, or asynchronously as a monitor, given that a second pass over every answer roughly doubles the cost of the system.
• How "the sources do not address this" is distinguished from "the sources contradict this", since the first is a retrieval problem and the second is a generation problem.

Failure modes

• An answer that correctly declines, saying the provided material does not cover the question, is scored as unsupported, so the gate penalises precisely the behaviour it exists to encourage.
• The evidence set recorded for validation is the candidate pool rather than the passages actually placed in the context, so the check is run against material the generator never saw.
• Claim decomposition splits a hedged sentence into a bare assertion and a discarded qualifier, and the assertion is then judged on its own and fails.
• An answer synthesising two passages is rejected because no single passage entails it, and the product loses the ability to answer anything requiring one step of inference.
• Passages exceed the validator's reliable range, and support decisions degrade as documents get longer rather than failing in any visible way.
• The threshold is set once from a pilot set and never revisited; retrieval improves, the supported proportion rises, and the gate stops binding without anyone observing it.
• Validation runs after the answer has streamed, so the only available remedy is a correction beneath text the user has already acted on.
• Arithmetic and aggregation are treated as claims to be entailed, no source states the total, and every computed answer fails for a reason that has nothing to do with grounding.
• The validator is given the question as well as the answer and starts judging whether the answer is good rather than whether it is supported, which is a different and much softer test.

Related capabilities

• Hallucination Detection — the check for fabrication where no evidence set exists to compare against
• Output Guardrails — the checkpoint where a failing verdict is acted on
• Human Approval Gates — where an answer that fails but matters is sent
• Citation and Grounding — the reader-facing attribution this gate is not
• Retrieval Evaluation — measures whether the evidence was found, offline, over a population

Applies: Production readiness · Design questions