How should AI answers be separated into atomic claims, and how should accuracy, completeness, recency, evidence and severity be scored?

Direct answer

An atomic claim is the smallest material proposition that can be assessed as true or false on its own. First split the answer into these claims. Then record for each the correct entity, accuracy, completeness, recency, source support, materiality and potential harm. Treat mention, citation and recommendation as separate events. A critical error cannot disappear into an average of many immaterial truths; NOMOS GEO calls this the non-compensation rule.

In plain language

The sentence 'The shop is open today, is in the city centre and offers free delivery' contains three claims. The first two may be true and the last false. The misleading delivery information does not become correct because most of the sentence is right.

Why this matters

One overall true-or-false label can conceal a material error. A high average score must not make a critical defect acceptable, such as a wrong identity, dangerous advice or unsupported authority.

Do not confuse

Mention means that an entity's name appears; it does not establish the accuracy of a claim.
Citation means that a source is named; it does not by itself prove that the source supports the relevant claim.
Accuracy is agreement with reference truth; completeness asks whether necessary information was omitted.
Recency asks whether the claim remains valid at the time of assessment.
Materiality or severity concerns effect on decisions and harm, not how noticeable an error appears.

What should you do?

Split the answer into atomic claims containing subjects, relationships, properties, numbers, dates, scopes and recommendations.
Link each claim to the correct entity and its corresponding record in the Truth Pack.
Assess accuracy, completeness, recency and evidential support in separate result fields.
Record mention, citation, recommendation and representation outcome without conflating them.
Classify severity through affected people, decision impact, reversibility and potential harm.
Apply the non-compensation gate to a critical error and do not close it through an overall average.
Preserve the judgement, rationale, evidence, auditor and uncertainty in the claim ledger.

How do you audit it?

Were several propositions scored as one claim?
Does every claim link to the correct entity and Truth Pack entry?
Were accuracy, completeness, recency and source support assessed separately?
Was the mere appearance of a citation counted automatically as evidential success?
Were favourable but false claims marked as errors?
Was a critical error compensated for by a high average score?
Were UNRESOLVED claims forced into a true or false class?

Limit

Atomic decomposition and severity classification involve human judgement. The NOMOS GEO thresholds and non-compensation rule are a proposed audit method; they are not an official common scoring system of NIST, every sector or applicable law.