Typed Evidence Checks Distinguish Citation Presence from Support in a Controlled Scientific Corpus

Abstract

This illustrative numerical study separates source existence, lexical relevance and proposition support in a controlled scientific style corpus. A seeded generator creates 240 fictional arithmetic records for eighty material lots and 560 claims spanning supported forms and changes to number, unit, condition, negation or entity. All supported evidence is retrieved first by a specified weighted cosine retriever. A citation presence rule accepts all 400 unsupported claims; a prescribed lexical threshold accepts 328, giving precision 32.8 percent. A six field parser and exact tuple check accept every supported claim and reject every unsupported claim within the declared closed grammar. Retrieving three passages adds no benefit over one. Removing source response content while retaining citation identity reduces supported recall to zero, although rejection of the majority class gives 71.4 percent accuracy. This is a symbolic attribution proxy with no language model, real scientific measurements or open language verification claim.