In Text-to‑SQL and AI‑agent analytics, successful SQL execution does not guarantee semantic correctness after business definitions change. This work investigates how to re‑validate fixed queries when refund attribution, tax treatment, eligible populations, or exchange‑rate dates are updated.
We introduce an executable verification profile that binds a revision request and knowledge‑cutoff to legally curated counterexample data and performs exact monetary comparisons. A preregistered experiment on 32 newly authored revision families and two local producer bundles retained 192 query‑definition cells: 144 non‑conforming, 36 finite‑suite agreeing, and 12 unresolved. Among the non‑conforming cells, 63 still agree on the current snapshot.
With an eight‑ledger budget, obligation‑guided checking uncovered 137 discrepancies versus 127 for a generic ordering, a paired gain of 6.9 percentage points (family bootstrap 95 % interval [2.1, 13.6]; exact family sign‑flip p = 0.0156). Every ledger arm rejected none of the 36 agreeing cells. Current‑obligation and equal‑weight scoring matched the guided result, indicating that coverage rather than revision‑specific priority drives the benefit.
Controlled assays, a natural‑query pilot, and repeated sampling each provide separate mechanism and robustness evidence. A legal discrepancy refutes overall conformance; agreement remains limited to the recorded suite. The study establishes an auditable business‑semantic revalidation profile within its authored schema and local‑model scope.
Review