Context decay in public repositories: an observational snapshot (v1)

A single observational snapshot, taken on 2026-10-18 through the GitHub REST API, of the agent context files (AGENTS.md, CLAUDE.md, GEMINI.md, .github/copilot-instructions.md, .cursorrules) in the most-starred public GitHub repositories, and of how often the repository paths they mention do not resolve in the current tree. It describes; it does not test a cause. This is the study package of the preprint arXiv:2610.31306: the protocol fixed before collection (in README.md and the header of collect.py), the collection, audit, summary and figure scripts (Apache License 2.0), the data the author derived (CC BY 4.0; third-party strings such as repository names and file paths keep their sources' terms), a codebook and SHA-256 checksums. One frame row and three candidate rows are withheld under a privacy list applied at collection and are still counted in the totals. Commit SHAs were not recorded, so the snapshot date and the default-branch names are the anchors. The v1 detector frozen for a registered replication, detect_v1.py, has SHA-256 6cc8c8ec0bff5fafcb6b2690ef21205124f456fcb717c6ceed836102107b758c; it was copied from the collect.py of this package, SHA-256 afd36688ac025217464f1dafd98698be9b605d6d7593288dcf3f88696d938640. Contact and removal requests: [e-mail].