{
  "_comment": "OFFLINE MIRROR of authoritative scholarly metadata, used in place of the live CrossRef / Semantic Scholar APIs which are rate-limited (429/503) inside the sandbox. This file is oracle-only (mounted at /oracle, never visible to the agent). It is a snapshot of REAL publication metadata about the *correct* entries in the field; it deliberately does NOT list which citations are fake. The oracle derives the fake citations by genuine computation: each bib entry is verified against this snapshot (DOI registrant-prefix lookup + fuzzy title match against the real-paper index), and any entry that fails to verify is flagged. The same elimination an online checker would perform, just sourced from a cached snapshot instead of a flaky network call.",
  "registered_doi_prefixes": {
    "_comment": "Real ISO 26324 DOI registrant codes (the digits after '10.') and the registration agency / publisher they belong to. A DOI whose registrant prefix is absent here cannot resolve at a real registry. Prefixes 1234 and 5678 are the canonical placeholder/example prefixes used in tutorials and are NOT assigned to any registrant.",
    "1038": "Springer Nature",
    "1126": "American Association for the Advancement of Science (Science)",
    "1016": "Elsevier",
    "1109": "IEEE",
    "1145": "Association for Computing Machinery (ACM)",
    "18653": "Association for Computational Linguistics (ACL Anthology)",
    "1007": "Springer",
    "1093": "Oxford University Press",
    "1073": "PNAS / National Academy of Sciences",
    "1056": "Massachusetts Medical Society (NEJM)",
    "1001": "American Medical Association (JAMA)",
    "1002": "Wiley",
    "1101": "Cold Spring Harbor Laboratory (bioRxiv)",
    "1186": "BioMed Central",
    "1371": "Public Library of Science (PLOS)",
    "1063": "AIP Publishing",
    "1103": "American Physical Society (APS)",
    "1021": "American Chemical Society (ACS)",
    "1086": "University of Chicago Press",
    "1075": "JSTOR / various",
    "48550": "arXiv (DataCite)",
    "5281": "Zenodo (DataCite)"
  },
  "real_papers": [
    {"title": "Highly Accurate Protein Structure Prediction with AlphaFold", "year": "2021", "venue": "Nature"},
    {"title": "LILA: A Unified Benchmark for Mathematical Reasoning", "year": "2022", "venue": "EMNLP"},
    {"title": "Molecular Structure of Nucleic Acids: A Structure for Deoxyribose Nucleic Acid", "year": "1953", "venue": "Nature"},
    {"title": "TriviaQA: A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehension", "year": "2017", "venue": "ACL"},
    {"title": "The New Frontier of Genome Engineering with CRISPR-Cas9", "year": "2014", "venue": "Science"},
    {"title": "CLUE: A Chinese Language Understanding Evaluation Benchmark", "year": "2020", "venue": "COLING"},
    {"title": "CMMLU: Measuring massive multitask language understanding in Chinese", "year": "2024", "venue": "ACL Findings"},
    {"title": "ChID: A Large-scale Chinese IDiom Dataset for Cloze Test", "year": "2019", "venue": "ACL"},
    {"title": "HellaSwag: Can a Machine Really Finish Your Sentence?", "year": "2019", "venue": "ACL"},
    {"title": "Molecular Biology of the Cell", "year": "2014", "venue": "Garland Science"},
    {"title": "Molecular Cloning: A Laboratory Manual", "year": "2001", "venue": "Cold Spring Harbor Laboratory Press"},
    {"title": "Attention is All You Need", "year": "2017", "venue": "NeurIPS"},
    {"title": "Deep Residual Learning for Image Recognition", "year": "2016", "venue": "CVPR"},
    {"title": "MCIP: Protecting MCP Safety via Model Contextual Integrity Protocol", "year": "2025", "venue": "EMNLP"},
    {"title": "Provably Secure Steganography", "year": "2009", "venue": "IEEE Trans. Computers"},
    {"title": "Adversarial watermarking transformer: Towards tracing text provenance with data hiding", "year": "2021", "venue": "IEEE S&P"},
    {"title": "Natural Language Watermarking via Paraphraser-based Lexical Substitution", "year": "2023", "venue": "Artificial Intelligence"},
    {"title": "Meteor: Cryptographically secure steganography for realistic distributions", "year": "2021", "venue": "ACM CCS"},
    {"title": "BoolQ: Exploring the Surprising Difficulty of Natural Yes/No Questions", "year": "2019", "venue": "NAACL"},
    {"title": "Efficiently Identifying Watermarked Segments in Mixed-Source Texts", "year": "2025", "venue": "ACL"},
    {"title": "AGENTVIGIL: Automatic Black-Box Red-teaming for Indirect Prompt Injection against LLM Agents", "year": "2025", "venue": "EMNLP Findings"},
    {"title": "Pre-trained Language Models Can be Fully Zero-Shot Learners", "year": "2023", "venue": "ACL"},
    {"title": "M-RewardBench: Evaluating Reward Models in Multilingual Settings", "year": "2025", "venue": "ACL"}
  ]
}