All docs
ConceptsClaims, grounding and where to correctdashboard

Claims, grounding and where to correct

How an answer is split into claims and checked against its sources and rows, what each verdict means, and where the fix belongs.

Claims and their verdicts

What it is. A claim is one fact in the answer. Each gets a verdict: Supported, Contradicted (a source says otherwise), Not in sources (nothing supports it), or, for Document analysis, Framing (not counted) and Follows from verified findings.

Why it matters. You see exactly which sentence is the problem, not just a low score.

How it relates. Claims drive Faithfulness, the Grounding and No contradictions checks, and where to correct.

Where you see it. An answer's Claims tab and Full grounding audit. Every case: Claims and figures.

Example. "The out-of-pocket maximum is $6,500" against "$6,000 individual": Contradicted.

Claim explanations

What it is. One plain sentence under Why for each claim that didn't pass. A contradicted claim shows both values and names the source; a claim not in the sources says what is missing and is never called wrong.

Why it matters. A reviewer can confirm or reject a verdict in seconds.

How it relates. The source is named by its source (else its id), so give documents readable names.

Where you see it. An answer's Claims tab.

Example. "The answer says $6,500; Summary of Benefits 2026, p.2 says $6,000." / "The sources don't give this figure ($900,000)."

Grounding and documents

What it is. Checking claims against what the answer was given: reference_materials, retriever outputs, tool results and query rows. Each passage is read with its document's title or source as context.

Why it matters. A passage that never repeats the plan name is still understood as part of that plan.

How it relates. Figures are compared by value ("$6,000", "USD 6,000"). A title never supplies a figure: it must be in the passage. With no sources, grounding isn't scored.

Where you see it. An answer's Claims and Scores tabs.

Example. "Acme's short-term disability benefits start after 7 days" is Supported by a passage from "Acme Short-Term Disability (STD) Plan 2026, section 3".

SQL answers checked against rows

What it is. When you send the rows your query returned, each figure in the answer is computed from them in code: a cell, a total, a count, an average or a share.

Why it matters. A correct total is proven correct and a misread one is caught, without a model guessing.

How it relates. A figure passes only when exactly one computation over the rows gives it; if two do, it is left unverified, and Max-Accuracy can say which is meant. variA/Bly never connects to your database.

Where you see it. An answer's Checks tab ("SUM(paid) over 3 rows = 875,000.50") and Claims tab.

Example. Rows North 425,000.50 / South 300,000 / West 150,000: "Total paid was $875,000.50" is Supported; "North paid $300,000" is Contradicted.

Where to correct

What it is. The step where the fix belongs: retrieval, a tool call, routing, a hand-off, the answer itself, or setup.

Why it matters. A low score tells you something is wrong; this tells you which prompt, retriever or contract to change.

How it relates. If a source you supplied contradicts the answer, the answer step is named (the right fact was retrieved). Retrieval is named only when the sources lack the fact and don't match the question. Setup means a check is Not configured.

Where you see it. An answer's Summary and Checks; Evaluation › Issues grouped by cause.

Example. "benefits-agent, writing the answer: the right reference was retrieved (relevance 0.82) but the answer states different numbers".