MTM-C (Machine Commands Machine) is the second interaction mode of the Machine Teaches Machine (MTM) framework established in Paper 2 of this series. Where MTM-T describes the transfer of confirmed linguistic knowledge from an LLM Teacher to a deterministic CGPA Student, MTM-C describes the downstream physical consequence of that transfer: the Student's translation of confirmed intent into physical-world device commands, executed asynchronously in under 150 milliseconds, without LLM involvement at runtime. MTM-C's defining architectural property is the structural decoupling of conversational latency from physical actuation latency. In CGPA, the verbal response fires from the Tier 1 pattern cache in under 20 milliseconds — before the device command is dispatched. The device command executes asynchronously in parallel via a protocol-agnostic device registry. These two latency budgets are co-designed, independent, and simultaneous. This architecture is structurally impossible in agent paradigms where both the verbal response and the device command require sequential LLM inference steps. MTM-C's most consequential property is its physical success gate: a device command is written to the pattern cache only when the physical device confirms successful execution. A command that fails — because the device is offline, out of range, or returns a protocol error — is never cached. The Student learns only from actions that demonstrably changed the physical world. This grounds the MTM-T learning loop in physical reality rather than linguistic plausibility, a property that is architecturally essential for safety-critical applications including healthcare device control and emergency escalation. This paper formally specifies the MTM-C actuation path, the device registry architecture, the asynchronous dispatch and success gate mechanism, the IoT protocol support matrix spanning six protocols across four deployment tiers, and the fail-safe behaviour model. MTM-C is demonstrated across three deployments: Botler (hotel room butler — KNX, Zigbee, WiFi, IR blaster, RS-232 across ten room scenes), Virtual Carer (elder care companion — BLE wearables, RS-232 wheelchair control, Android telephony emergency escalation), and MAOI (FnB POS — kitchen display routing, cash drawer peripheral actuation). A systematic comparison with agent-based IoT architectures establishes that MTM-C's success-gated, LLM-free actuation path cannot be replicated within the agent paradigm without abandoning the agent's core architectural property: LLM inference as the universal mediator of all decisions.
TRUONG VIET PHAN (Tue,) studied this question.