A transaction control layer for governed agentic commerce: reference architecture implementation and controlled evaluation

Abstract

A weak answer from a shopping assistant may inconvenience a customer. A weak action from a commerce agent can move money, reserve inventory, change a delivery commitment, or issue a refund. Once an agent can create those effects, the central question is no longer whether its response sounds correct. The question is whether the action is still authorized and commercially valid when the enterprise system commits it. This study introduces a Transaction Control Layer (TCL) that separates an agent's proposal from permission to execute. TCL binds each approved action to grounded intent, delegated authority, current authoritative state, a short validity period, and a stable idempotency key. It then records and reconciles the resulting side effects. We formalize commit validity across five dimensions and define five operational decisions: allow, modify, hold, deny, and escalate. The design is examined with CAV-Bench v1.0.0, an Apache-2.0 reference artifact containing 40 synthetic scenarios and five deterministic architecture profiles. The direct profile reaches an outcome success rate of 0.925, yet only 0.250 of its runs are commit-valid. Adding policy checks, commit-time guards, reconciliation, and lifecycle recovery raises commit-valid success to 0.500, 0.750, 0.875, and 1.000. These values show that the benchmark responds to the controls it was designed to test. They do not estimate real model failure rates or establish production effectiveness. An external replication protocol is therefore included as the next step toward independent validation.