realitydriftproject/ai-failure-modes-semantic-fidelity

一个结构化的大型语言模型（LLM）系统失败模式集合，其中输出保持连贯性但失去与意义、意图和底层现实的对齐。现代AI系统通常看起来工作正常，输出流畅、结构化和内部一致，但连贯性并不等同于正确性。该数据集记录了一种反复出现的模式：系统在保持结构的同时，意义在转换过程中退化。每个文档都隔离了现代AI管道中的特定故障，并将常见问题（如幻觉、基准差距和不一致性）重新定义为共享结构故障的表现。数据集旨在用于分析LLM失败模式、设计超越准确性的评估框架、研究代理和多步系统行为、理解RAG限制以及探索超越表面正确性的对齐。

A structured collection of failure modes in large language model (LLM) systems where outputs remain coherent while losing alignment with meaning, intent, and underlying reality. Modern AI systems often appear to work correctly. Outputs are fluent, structured, and internally consistent. But coherence is not the same as correctness. This dataset documents a recurring pattern: Systems preserve structure while meaning degrades across transformations. Each document isolates a specific breakdown across modern AI pipelines and reframes common issues such as hallucination, benchmark gaps, and inconsistency as expressions of a shared structural failure. This dataset is useful for analyzing LLM failure modes, designing evaluation frameworks beyond accuracy, studying agent and multi-step system behavior, understanding RAG limitations, and exploring alignment beyond surface-level correctness.