PRISM

This replication package supports the paper "Signal-Conditioned Prompting for LLM-Assisted Requirements Inspection". It contains two components: the PRISM signal computation pipeline and the PRISM-Copilot proof-of-concept experiment. Component 1 — PRISM Signal Pipeline (data/, validation/, anchors/) PRISM computes three requirement-level readiness signals — Structural Criticality (C), Linguistic Specificity (S), and Volatility Exposure (V) — over a 262-requirement corpus derived from the teraPROMISE NFR e-commerce subset. The pipeline runs in five reproducible steps: corpus cleaning (Step 1), sentence embedding (Step 2), redundancy resolution at τ = 0.86 (Step 3), signal scoring (Step 4), and Pareto front structuring (Step 5). All derived artifacts, embeddings, signal scores, and Pareto front assignments are provided under data/derived/. Validation experiments covering construct validity, behavioral plausibility, signal independence, and robustness under noise perturbation are reproducible via validation/run_validation.sh. Linguistic anchors used in S and V construction are externalized under anchors/ for transparency. Component 2 — PRISM-Copilot Experiment (experiments/prism_copilot_poc/) PRISM-Copilot instantiates the Signal-Conditioned Inspection Prompting (SCIP) pattern. The experiment evaluates 50 stratified requirements under three prompting conditions: no guidance (Condition A), signal-conditioned (Condition B), and generic heuristic (Condition C). The package includes the corpus master file with pre-computed signal vectors, the stratified evaluation sample, all three prompt templates, raw and parsed LLM outputs (GPT-4o, temperature 0), practitioner review sheets and responses from five independent reviewers, inter-rater agreement data, recall alignment results, and all analysis scripts (Steps 01–11). The signal-to-policy mapping protocol is documented in experiments/prism_copilot_poc/protocol/signal_policy_mapping.md. Reproducibility All paths are repository-relative. No absolute paths are used. Python 3.9+ is required with dependencies: numpy, pandas, scipy, networkx, scikit-learn. This package is anonymized for double-blind review. Author identifiers will be added upon acceptance.本复现包支持论文《面向大语言模型辅助需求审查的信号条件提示方法》。 本包包含两大组件：PRISM信号计算流水线与PRISM-Copilot概念验证实验。 组件1 — PRISM信号流水线（对应目录：data/、validation/、anchors/） PRISM可基于源自teraPROMISE非功能需求（Non-Functional Requirements，简称NFR）电商子集的262条需求语料库，计算三种需求级就绪信号：结构关键性（C）、语言特异性（S）与波动性暴露度（V）。该流水线包含五个可复现执行步骤：语料清洗（步骤1）、句子嵌入（步骤2）、τ=0.86阈值下的冗余消解（步骤3）、信号评分（步骤4）与帕累托前沿构建（步骤5）。所有衍生产物、嵌入向量、信号评分及帕累托前沿分配结果均存放于data/derived/目录下。针对结构效度、行为合理性、信号独立性及噪声扰动下鲁棒性的验证实验，可通过validation/run_validation.sh脚本完成复现。用于构建S与V信号的语言锚点已外置至anchors/目录，以提升研究透明度。 组件2 — PRISM-Copilot实验（对应目录：experiments/prism_copilot_poc/） PRISM-Copilot实现了信号条件化审查提示（Signal-Conditioned Inspection Prompting，简称SCIP）范式。本实验在三种提示条件下对50条分层抽样需求开展评估：无引导条件（条件A）、信号条件化提示（条件B）与通用启发式提示（条件C）。本复现包包含以下内容：带预计算信号向量的语料主文件、分层评估样本、全部三种提示模板、原始与解析后的大语言模型（LLM）输出（GPT-4o，温度参数设为0）、5名独立评审员的从业者评审表与回复、评分者间一致性数据、召回率对齐结果，以及所有分析脚本（步骤01至11）。信号至策略映射协议已在experiments/prism_copilot_poc/protocol/signal_policy_mapping.md文件中予以说明。 可复现性说明 所有路径均为仓库相对路径，未使用绝对路径。本项目需Python 3.9及以上版本，依赖库包括numpy、pandas、scipy、networkx与scikit-learn。本复现包已做匿名处理以支持双盲评审，作者信息将在论文录用后补充。