Ornella Bahidika and Joel Allou explain why a language model should not have unrestricted authority over a tutoring workflow. Their voice-tutor harness defines lesson states, including teaching, checking understanding, grading and advancing, rather than asking the model to manage the whole lesson implicitly.
Ornella Bahidika and Joel Allou assign bounded conversational tasks to the model and validate its outputs before changing state. The demonstration shows lesson queues, logs and transitions. They discuss using a smaller model for responsive interaction, without claiming that model choice alone establishes reliability.
Ornella Bahidika and Joel Allou extend the same control pattern to onboarding, operations and coding workflows. The central distinction is between proposing an action and granting authority to execute it. The demonstration illustrates that separation, not a general safety guarantee.
Watch on YouTube
