TR-2026-02: Transactional State Reconciliation, Edge Fleet Custody, and ISA-95/Purdue Safety Boundaries for Autonomous Industrial Operations

Deploying autonomous AI agents across discrete and process manufacturing environments introduces acute cyber-physical risks when enterprise cloud control planes interact with plant-floor systems across ERP, MES, WMS, and QMS. Extending the self-optimizing factory paradigm formulated in Durga (2025, DOI: 10.38124/ijsrmt.v4i8.1046) and the cloud-native platform architectures established in Durga (2025, DOIs: 10.38124/ijsrmt.v4i7.1045, 10.38124/ijsrmt.v4i6.1044, 10.38124/ijsrmt.v4i5.1043, and 10.5281/zenodo.23723284), this technical report presents deterministic benchmarks for governed industrial agent operations and edge AI model fleets. We evaluate a 4-mode Pre-Release Recovery Reservation Lease with compare-and-swap (CAS) fencing, a cycle-safe transitive rollback DAG across ISA-95 Level 3/4 systems, a hard physical safety firewall isolating Level 0/1 Programmable Logic Controllers (PLCs) and Safety Instrumented Systems (SIS), and a Two-Layer JCS/CBOR Merkle edge model custody protocol achieving sub-100 ms last-known-good (LKG) rollback under intermittent plant connectivity.