Sobriété Numérique and the Frugal Tutor: Environmental Justice, Accessibility, and the Cognitive Floor of Edge-Deployed Language Tutoring

Abstract

BMO is an offline, voice-to-voice French tutoring agent designed to run entirely on consumer CPU hardware, built to test how far AI language tutoring can move away from centralized cloud APIs toward local, sovereign computation. This paper asks a specific empirical question: how small can a language model be shrunk before it loses its ability to teach effectively? Three configurations, spanning 3.4B to 7.6B parameters, are evaluated on identical hardware for teaching quality, response naturalness, energy per turn, and latency, using a documented LLM-as-judge protocol adopted after an automated scoring method was shown to have no real discriminative power. The paper introduces the Sovereign Utility Score, an original metric combining teaching quality and resource cost into a single number, and reports its central finding: the cheapest, fastest configuration is also the one most likely to silently confirm a learner's grammatical error as correct, while the best teacher costs roughly twice as much to run. Framed through the French concept of sobriété numérique (digital sobriety), the paper connects this trade-off to environmental justice, accessibility, and linguistic-justice concerns beyond pure computational efficiency.