文章背景与核心概要
在大型金融机构中，传统的IT变更管理通常依赖主观的手动问卷，这些问卷不仅容易被人为操纵，而且在识别高风险变更方面效果不佳。本文介绍的 SENTRY 是一个全新的风险评估平台，它利用确定性的机器学习流水线取代了这些低效的手动流程。
SENTRY 通过将结构化运营元数据、应用程序依赖图谱和历史事件数据与混合 RAG（检索增强生成）方法相结合，为传统系统提供了一种可扩展、可解释且高准确率的替代方案。评估结果表明，SENTRY 的 ROC AUC 达到了 0.87，整体准确率为 85%，高风险变更的检测效率约为传统流程的 3.25 倍。
SENTRY: Deterministic, Intelligent Risk Assessment for IT Change Management
arXiv: 2608.21203
Date: August 21, 2026
Authors: Daniel Arulpragasam, Christer Henrysson, Ella Ly, Deepika Anbalagan, Leo Feng
Summary
在大型金融机构中，传统的IT变更管理通常依赖于主观、手动填写的问卷，这些问卷容易遭到人为操纵，且在识别高风险变更方面效率低下。SENTRY 是一个全新的风险评估平台，它用确定性的机器学习流水线取代了这些手动流程。通过将结构化运营元数据、应用程序依赖图谱和历史事件数据与混合 RAG（检索增强生成）方法相结合，SENTRY 为传统系统提供了一个可扩展、可解释且高度准确的替代方案。
Summary
In large financial institutions, traditional IT change management often relies on subjective, manual questionnaires that are prone to manipulation and ineffective at identifying high-risk changes. SENTRY is a new risk assessment platform that replaces these manual processes with a deterministic machine learning pipeline. By integrating structured operational metadata, application dependency graphs, and historical incident data with a hybrid RAG (Retrieval-Augmented Generation) approach, SENTRY provides a scalable, explainable, and highly accurate alternative to legacy systems.
The Problem: Limitations of Manual Questionnaires
当前行业标准的风险评估方法经常受到以下局限性的困扰： * 主观性： 过度依赖自我申报的数据，导致评分不一致。 * 博弈性： 问卷极易被操纵，无法区分例行更新和高影响力的变更。 * 缺乏洞察： 现有流程难以将历史事件数据与当前的变更请求进行关联。
The Problem: Limitations of Manual Questionnaires
Current industry standards for risk assessment are frequently hampered by: * Subjectivity: Reliance on self-reported data leads to inconsistent scoring. * Gaming: Questionnaires are easily manipulated, failing to distinguish between routine updates and high-impact changes. * Lack of Insight: Existing processes struggle to correlate historical incident data with current change requests.
The Solution: The SENTRY Architecture
SENTRY 引入了一个专为高风险、受监管环境设计的强健、确定性流水线：
- 混合 RAG 流水线： 结合语义检索与词法检索，从变更请求的非结构化文本中提取风险信号。
- 特征工程： 将复杂的文本风险信号压缩为单一的标量特征，用于模型输入。
- 模型核心： 采用梯度提升决策树（XGBoost）来维持确定性的决策过程。
- 可解释性： 结合 SHAP（SHapley Additive exPlanations）值，确保每一次风险预测都具备可审计性和透明度。
The Solution: The SENTRY Architecture
SENTRY introduces a robust, deterministic pipeline designed for high-stakes, regulated environments:
- Hybrid RAG Pipeline: Combines semantic and lexical search to extract risk signals from unstructured text in change requests.
- Feature Engineering: Compresses complex textual risk signals into a single scalar feature for model input.
- Model Core: Utilizes Gradient-Boosted Decision Trees (XGBoost) to maintain a deterministic decision-making process.
- Explainability: Incorporates SHAP (SHapley Additive exPlanations) values to ensure every risk prediction is auditable and transparent.
Performance Metrics
在企业级规模的数据上进行评估时，SENTRY 展现出相比现有方法的大幅性能提升： * ROC AUC： 0.87 * 总体准确率： 85% * 检测效率： 识别高风险变更的效率约为传统流程的 3.25 倍。
Performance Metrics
Evaluated against enterprise-scale data, SENTRY demonstrates significant improvements over existing methodologies: * ROC AUC: 0.87 * Overall Accuracy: 85% * Detection Efficiency: Identifies high-risk changes at a rate approximately 3.25 times higher than the legacy process.
Access & Resources
Access & Resources
