Sentencing decision-making is the closest aspect of criminal justice to the substantive outcome of a decision, requiring strict compliance with the boundaries of statutory penalties, as well as careful differentiation between aggravating, mitigating, lenient and stringent circumstances. Although the existing judicial AI research has accumulated a lot of results in legal judgment prediction, case search and legal model evaluation, the “accuracy” in the public discussion is mostly compressed into a single classification accuracy, F1 value or regression error, which is difficult to correspond to the high-risk sentencing assistance scenarios directly. In this paper, we define the accuracy of AI-assisted sentencing as five inter-coupled dimensions: legal element alignment, sentencing interval accuracy, dynamic robustness, rationale traceability, and deployment security. In the absence of an online validation interface for public courts, this paper incorporates six types of public benchmarks, 18 representative studies, and four governance documents, constructs a path-level evidence corpus, scores the six types of technological routes with normalized scores, and proposes a closed-loop optimization path consisting of legal knowledge grounding, case-like retrieval, interval constraints, stepwise calibration, and uncertainty gating. The results show that the combined accuracy score of the coupled path of validation correction and uncertainty gating is 79.3, which is higher than that of the directly prompted large model (58.5) and the static supervised model (54.0); the robustness on the law change and innocence determination scenarios is improved by 27.1 and 28.6 points, respectively; and the serious error rate can be compressed down to 8.9% with 63% automatic coverage. The study shows that the key to a sentencing support system is not the optimization of a single metric, but the synchronization of multidimensional accuracy, risk boundaries, and manual review interfaces. For practical deployment, a more feasible path would be to position AI as a “candidate sentencing range generator, rationale and case presenter, and risk warning device”, while leaving the final sentencing discretion firmly with the judge.
