We develop a Bayesian source reconstruction framework with a constant-release assumption for atmospheric radionuclide releases, comparing the performance of log-Gaussian, log-Cauchy, and log-Laplace likelihood functions. This framework is validated using two representative cases: the first release of European Tracer Experiment (ETEX-I) and the 2017 106Ru event. For ETEX-I, the log-Gaussian likelihood achieves the best performance, with a source location error of 29.854 km and a total release error of 4.507%, producing narrow a posteriori distributions closely matching the true source. For the 2017 106Ru event, the log-Laplace likelihood performs best, with a location error of 96.511 km and a total release amount consistent with the reported value, though it estimates a longer release period than reported. Our results reveal that reconstruction performance aligns with the likelihood’s ability to fit observation–simulation discrepancies. To facilitate likelihood selection, we propose a Bayesian a posteriori evaluation criterion assessing both robustness to outliers and sensitivity to scale parameters, systematically identifying the optimal likelihood under the employed source–receptor data. This framework reliably reconstructs source parameters with quantified uncertainties, representing a significant advancement for nuclear emergency response. Its data-driven approach eliminates manual parameter selection, enhancing its practicality and applicability in real-world scenarios.