A Layered Audit Protocol for AI-Assisted Synthesis of Research Frontiers

Data-and-code deposit accompanying the technical report A Layered Audit Protocol for AI-Assisted Synthesis of Research Frontiers. It contains 489 coded open-problem statements (limitation and future-work sentences) extracted from 96 machine-readable papers of a curated AI corpus, the coding schema and its two sharpenings, inter-coder reliability (Cohen's kappa v1 and post-sharpening), the reconciled gold standard, the frontier map (six families, nineteen sub-frontiers), six blind-discovered derived frontiers plus one analyst-nominated applied frontier, 32 formalised research questions, 98 deterministically verified off-corpus candidate papers with relevance verdicts, and the full pipeline scripts and agent prompts. The deposit contains only short verbatim excerpts (the open-problem statements) and metadata; full text of the reviewed papers is not redistributed, consistent with third-party copyright. The central finding is methodological: quote extraction was faithful and coding reproducible, but unaudited aggregation over-claimed by roughly two- to threefold, so every synthesis layer must be adversarially audited. Code is licensed MIT; data, codebooks, and method note are CC BY 4.0. AI-assistance disclosure and agent prompts are included (APPENDIX_G_agent_protocol.md).