Machine Teaches Machine (MTM-T): A Framework for Cross-Architectural AI Knowledge Transfer

Machine Teaches Machine (MTM-T) is a theoretical framework describing an architectural pattern in which a generative AI system — the LLM Teacher — resolves novel operational tasks through semantic inference and transfers the confirmed outcomes of successful real-world transactions to a deterministic, non-neural system — the CGPA Student — which stores and reapplies those resolutions permanently. The Teacher teaches once per confirmed surface-form variant of a pattern class. The Student remembers permanently. Once a pattern is transferred, the Student resolves that class of task independently at under 20 milliseconds; the Teacher is retired from that class. MTM-T operates within a three-tier resolution cascade: a pattern cache (Tier 1, <20ms), a playbook-driven grammar and regex parser (Tier 2, <30ms), and an LLM fallback (Tier 3, <500ms). Confirmation (yes/no) and slot-fill responses (numbers, item names) are handled as a probe-response branch within the Tier 2 state machine — they are not a separate tier. A scene scan path (D-616) activates after all tiers miss, routing utterances dynamically to the correct scene. This paper formally specifies MTM-T, the three-tier cascade, the Tier 2 compile/distill architecture and its runtime regex mechanism, the federated cache seeding model, the relational memory model, the probe-response handler, and the scene scan extension. It distinguishes MTM-T from Knowledge Distillation, Transfer Learning, Machine Teaching, and Federated Learning.