ADAM Methane-Monitoring Evaluation Dataset: Labeled Exposure Trials (D1) and Live Deployment Traces (D2) from a Four-Node Edge Multi-Agent Testbed

This dataset supports the evaluation of ADAM (Agentic Decentralized Autonomous Machines), a crew-based multi-agent framework for accountable edge intelligence in Decentralized Physical Infrastructure Networks (DePINs), evaluated on a methane-monitoring testbed of four Raspberry Pi 5 nodes with MQ-4 sensors, a co-located NDIR reference analyzer, on-device Gemma 3 1B (INT4) inference through Ollama, Weaviate semantic memory, and Fides Innova Proof-of-Authority governance. The workbook contains 18 documented sheets spanning two datasets and their derived records. D1 comprises 2,000 labeled exposure events (10 trials of 200; 900 anomaly, 1,100 normal), with ground-truth labels derived solely from the reference analyzer, scored under two evaluation modes: a full-pipeline benchmark in which nine systems classify identical events, and a trigger-gated run reproducing deployment semantics. D2 comprises 459 live coordination events from a 72-hour deployment, each with six per-stage latencies, resource counters, and trace-persistence flags, alongside 909 resource-measurement windows. Additional sheets record node-scaling runs (physical hardware at 1–4 nodes and a validated software scale-out model to 16, labeled per row), fitted scaling models, and security stress tests covering sensor injection, semantic-store poisoning, induced model failure, and external data egress. Every figure and table in the accompanying article recomputes from these sheets. The reference implementation, including automated parity checks against this workbook, is available at https://github.com/CryptoGuy1/ADAM. All values are static; formula cells were frozen to their computed values for archival stability. See the 00_README sheet for the sheet inventory, evidence classes, and authoritative denominators.