A readable explanation needs a path back to the data it describes. For a numeric claim, that path includes the returned value, the relevant unit, and the query context.
The numeric grounding component compares numbers in an answer with returned result values and associated facts. It can report matched, unmatched, derived and misattributed values. Its interface also carries row counts, column labels and query context. This is a bounded verification mechanism, not a certificate for every sentence the model produces.
Audit records provide another part of the explanation. They connect the question with the generated SQL, validation and approval state, and execution updates. The approval path consults this context before allowing work to proceed. Some context fields are configuration-dependent; a record should not be treated as proof that every optional check ran.
Together, these mechanisms make specific questions possible: Which data supported this number? Which query ran? What permission checks applied? What remained unresolved? A complete evaluation still needs to examine the intended meaning, missing results and unsupported claims. Fluency alone answers none of those questions.