Supplementary Materials for "An Architectural Pattern for Conversational and Agentic BPM with Model Context Protocol, Generative AI, and Defense in Depth: A Validated Instance and Formative Study"

Supplementary materials supporting the IEEE Access manuscript on a validated instance of an architectural pattern for AI-Augmented Business Process Management Systems (ABPMS). Materials are organized in nine logical packages (A–I) covering: component benchmarks across six LLM models (A); stochastic ablation of ten defensive layers (B); LLM-real multi-provider ablation of pre-LLM layers L1+L2 across five models in two capability tiers — tier small/fast (Anthropic Claude Haiku 4.5, OpenAI GPT-4o-mini, Google Gemini 2.5 Flash) and tier large (Anthropic Claude Sonnet 4, OpenAI GPT-4o) — totalling 1,000 LLM calls (C); formative study with N=25 lay users recruited via Prolific BR — anonymized data, briefing, and Apps Script form (D); statistical analysis including bootstrap BCa confidence intervals, Kruskal–Wallis omnibus test with exact permutation, MDE sensitivity analysis under observed dispersion, and robustness analyses (E); production engineering metrics during the pilot window (F); operational characterization of the 54 process versions generated by participants (G); per-agent layer mapping (H); and the detailed protocol of the structured state-of-the-art mapping including search strings, PRISMA-style flow, the I1–I4 inclusion criteria and E1–E4 exclusion criteria table, and the extraction template (I). All Prolific participant identifiers have been replaced with sequential pseudonyms P01..P45 (the 45 unique respondents effectively represented in the deposited dataset); the internal mapping is retained by the authors under documented request for audit purposes.本补充材料配套于发表于《IEEE Access》期刊的一篇学术论文，该论文针对人工智能增强型业务流程管理系统（AI-Augmented Business Process Management Systems，ABPMS）的架构模式验证实例展开研究。本补充材料按9个逻辑模块（A至I）进行组织，涵盖以下内容：六款大语言模型（Large Language Model，LLM）的组件基准测试（模块A）；针对10个防御层的随机消融实验（模块B）；针对预大语言模型层L1+L2的多厂商真实大语言模型消融实验，共覆盖两类能力层级下的5款模型：小型/快速层级（Anthropic Claude Haiku 4.5、OpenAI GPT-4o-mini、Google Gemini 2.5 Flash）与大型层级（Anthropic Claude Sonnet 4、OpenAI GPT-4o），总计开展1000次大语言模型调用（模块C）；通过Prolific BR平台招募的N=25名普通用户参与的形成性研究，包含匿名化数据、研究说明文档与Apps Script表单（模块D）；统计分析内容包括偏差校正加速（BCa）Bootstrap置信区间、带精确置换的克鲁斯卡尔-沃利斯整体检验、基于观测离散度的最小可检测效应（Minimum Detectable Effect，MDE）敏感性分析，以及稳健性分析（模块E）；试点运行阶段的生产工程指标（模块F）；参与者生成的54个流程版本的运行特征分析（模块G）；单智能体层映射关系（模块H）；结构化现有研究综述的详细方案，包含检索式、PRISMA式流程图、I1-I4纳入标准与E1-E4排除标准表格，以及数据提取模板（模块I）。所有Prolific平台参与者的标识符均已替换为连续化名P01至P45（本存档数据集共包含45名唯一受访者）；作者留存了内部映射关系，可根据书面申请用于审计用途。