FreshState: A Prospective Benchmark and Resource for Stale Evidence in Web-Augmented LLMs

FreshState is a benchmark and resource for detecting and verifying stale evidence in web-augmented information retrieval and retrieval-augmented generation pipelines. This artifact accompanies the paper “FreshState: A Prospective Benchmark and Resource for Stale Evidence in Web-Augmented LLMs.” It contains: - 696 high-confidence change-event seeds across three domains: Craigslist rental listings, GitHub releases, and PyPI packages;- a 602-example balanced, age-matched stale/fresh evaluation set for stale-evidence detection;- a naive-construction ablation set demonstrating the need for age matching;- collection, extraction, monitoring, validation, baseline, and reproduction code;- saved baseline, LLM verifier, and snippet-swap outputs;- extractor validation logs;- schema documentation and reviewer-facing reproduction instructions. The released artifact reproduces the paper’s reported tables from a clean checkout using the included scripts and cached outputs. The benchmark is intended to support stale-evidence detection, freshness-aware reranking, and evaluation of RAG pipelines under temporally invalid retrieved evidence. Project repository: https://github.com/kelu-look/freshstate Licensing:- Project-authored source code and reproduction scripts are licensed under the MIT License.- FreshState-curated benchmark records, evaluation sets, state and metadata files, validation logs, result files, prompt templates, saved model-output records included as experimental artifacts, schema documentation, and README documentation are licensed under Creative Commons Attribution 4.0 International (CC BY 4.0).- Underlying third-party source content, webpages, repositories, registry/platform content, trademarks, and third-party dependencies are not relicensed by this release.