Multi-Agent Behavioral Authentication Benchmark and Reproducibility Artifact

This artifact contains the finalized synthetic benchmark corpus, derived decision windows, validation manifests, result artifacts, curated replay code, Colab/GPU result archives, Raspberry Pi 5 profiling outputs, human quality-audit materials, and manuscript source/PDF for the accompanying paper on edge-oriented behavioral authentication in multi-agent LLM systems.

The benchmark formalizes behavioral authentication as rolling-window verification over multi-agent conversational traces. The released corpus contains 10,000 validated synthetic conversations and 19,216 labeled 5-turn decision windows spanning eight enterprise workflow families, four agent roles, and seven attack or condition types. Labels support binary verification of whether observed behavioral evidence in a short conversational window is consistent with the claimed agent identity.

The package is designed for reproducible evaluation replay rather than bit-identical regeneration of the synthetic conversations. The archived corpus is the citable benchmark object. The package includes deterministic validation summaries, model result artifacts, TF-IDF and neural baseline outputs, cross-generator transfer artifacts, RP5 edge-efficiency profiles, and a two-rater human quality-grader audit.