Below the Waterline: replication package for a multivocal review of agentic AI security, shadow AI governance, and LLM supply-chain risk

<p>Complete replication package for the multivocal literature review <em>Below the Waterline: A Survey of Agentic AI Security, Shadow AI Governance, and LLM Supply-Chain Risk in the Enterprise</em> (submitted to <em>Computers &amp; Security</em>, 2026).</p><p>Contains the full search protocol and the four screening and coding codebooks; every retrieved record (7,254 raw, 5,280 after deduplication) with the raw per-query API responses; the screening and eligibility decision for every record with its recorded exclusion reason; the coded corpus of 2,136 included works with evidence tier, failure primitive, sub-category, application stratum, role, and cross-cutting flag; the independent second-screener and second-coder samples from which the reported agreement statistics are computed; and the analysis and figure-generation scripts.</p><p>A single command, <code>python3 05_scripts/reproduce_all.py</code>, regenerates all 47 quantities reported in the paper from the released decision files, including the PRISMA counts, Cohen's kappa for screening (0.86, n=898) and for primary-primitive coding (0.89, n=427), the primitive-by-stratum gap map, and the evidence-tier sensitivity analysis. The script requires only the Python 3 standard library and exits non-zero if any number fails to reproduce.</p><p>Retrieval, screening, and coding were executed by generative-AI screeners operating under the written codebooks included here, as disclosed in the paper. This package is released so that those judgments can be audited rather than taken on trust.</p>