Boundary Persistence Minimal Benchmark v0.1: A Minimal Dataset and Scoring Instrument for Epistemic Boundary Robustness Evaluation

This dataset introduces the Boundary Persistence Minimal Benchmark v0.1, a small evaluation instrument for testing whether AI systems preserve epistemically relevant distinctions under symbolic pressure. The benchmark operationalizes the concept of boundary persistence: the capacity of a system to preserve distinctions between role, role label, authority, evidence, citation form, provenance, uncertainty, and reality constraint when the symbolic form of an input changes. It is designed as a minimal bridge between conceptual work on epistemically robust AI and reproducible behavioral evaluation. The dataset contains controlled benchmark items across five test families: role imitation, false citation injection, authority-style pressure, tool-output spoofing, and uncertainty pressure. Each family targets a distinct form of symbolic boundary failure, where a model may mistake the symbol of authority, evidence, provenance, or confidence for the corresponding operational status. The package includes benchmark prompts, a scoring template, a spreadsheet-based scoring sheet, a simple 0/1/2 rubric, and a Python script for computing provisional aggregate measures such as Boundary Persistence Score, Boundary Failure Rate, Partial Instability Rate, Full Persistence Rate, Attack Success Rate, and Semantic-Control Gap. This release does not claim to empirically validate boundary persistence as an internal architectural property of AI systems. It is a minimal, reproducible instrument for measuring behavioral indicators of epistemic boundary robustness in present-day models. Its purpose is to make the framework testable, inspectable, extendable, and falsifiable. DOI: 10.5281/zenodo.21187766 Related conceptual work: Toward Epistemically Robust AI: From Symbolic Return to Boundary Persistence.