What Can We Still Ask? Auditing Meaning Loss in Self-Report Data Pipelines

What Can We Still Ask? Auditing Meaning Loss in Self-Report Data Pipelines Two people can report opposite context–rating assignments yet become indistinguishable after export. An “it depends” option cannot prevent this if later transformations discard the association between a condition and its answer. We present a reproducible computational case study of task-relative auditing for self-report data pipelines. The audit separates whether a question is recoverable, whether its interpretation follows a declared version policy, and whether disclosure permits it to be answered. Building on established query determinacy and incomplete-information semantics, a machine-readable contract specifies questions, clarification vocabularies, interpretation rules, and appropriate failure witnesses. Exhaustive enumeration of 25 two-context and 125 three-context records produces 18 projection–query audits: ranges and rating multisets preserve some questions while losing context-specific answers. Further constructions show that even an unchanged partial record can lose exactness or acquire unwarranted certainty through reinterpretation, without an input collision. Version-aware examples preserve unselected reasons as recorded absences and decline to invent clarifications for unknown contracts. Against an independently documented TRAJEXA compatibility policy, nine synthetic component comparisons expose two version-2 interpretation violations. An implementation audit also identifies retained reason codes, optional prose disclosure, and unresolved annotation bindings. These are structural and software findings, not evidence of improved psychological measurement or human outcomes. The contribution is a bounded, executable case study: downstream answers should be constrained by the distinctions, interpretation rules, and disclosure authority that remain available. This deposit is a methods and computational case-study preprint, not a validated psychometric instrument. It applies established query determinacy and incomplete-information semantics; independent novelty and human benefit have not been established. The supplement contains executable constructed-data audits, the declared audit contract, a frozen dependency snapshot, source manuscripts and review/revision evidence. The version-policy audit retains two contract violations among nine synthetic comparisons. The English PDF is the main manuscript; the Japanese PDF is the corresponding full-text reading version. This deposit has not been peer reviewed.