Computational replication does not establish statistical validity when historical inputs may be misdated or mismeasured. This study evaluates how record-level data discre-pancies interact with lag specification and finite-sample inference in a three-variable vector autoregression (VAR). Using 30 annual observations supplied for Malaysia (1995–2024), we exactly reproduce the previously reported VAR(2) credit-equation coef-ficients, but identify a 2019 domestic-credit entry of 12.67% of GDP against a World Bank–derived historical comparator of 120.71%. A one-record scenario reverses the signs of both lagged takaful coefficients and changes their joint-exclusion p-value from 0.842 to 0.193 in VAR(2). A separate source audit finds that multiple takaful values correspond to published figures for other years or contribution categories: the supplied value for 2023 is RM15.64 billion, whereas the Malaysian Takaful Association reports 2023 total net con-tributions of RM17.3 billion. Across 12 prespecified data, transformation, and lag speci-fications, no conventional joint takaful exclusion test rejects at the 5% level. In calibrated 2,000-replication Monte Carlo experiments with 28 transformed observations, a log-level recording error of the observed magnitude reduces alternative-model rejection from 69.3% to 7.7% and raises null-model rejection from 3.8% to 6.4% within that design. These results demonstrate that exact numerical reproducibility can coexist with severe source-data and inference sensitivity. The contribution is a fully auditable replication, source-comparison, and contamination-stress-test case, not a causal estimate of takaful effects on credit.