The Mentor

The actor

The system has one actor: the Mentor. Everything else serves it. The Mentor is not a chatbot that answers questions — it is a counterpart that knows the learner and accompanies them through explanation and silence, encouragement and provocation, patience and demand.

What the Mentor generates is the learning experience itself: content, tone, timing, provocation, context, the sense of being accompanied. When a learner says "I want to understand why my bridge collapsed," no material exists that answers this question for this person in this state of mind. The Mentor creates it.

Three tiers of generation

Three tiers operate at different timescales. The first is real-time dialogue: Socratic exchange, tasks and feedback, generated in the moment. The second is simulation environments — a specialised Forge-AI asynchronously generates interactive physics labs, chemistry environments and engineering scenarios, and places them in a shared repository. Google DeepMind's Genie 3 has demonstrated that such generation is technically feasible; what it does not yet guarantee is the physical precision education demands. The third tier is long-lived deterministic engines — physics, chemistry — that compute the actual natural laws from lightweight instruction sets.

The combination matters. Tier-2 Forge-AI generates the environment and its pedagogical framing; Tier-3 engines compute the laws inside it. The generation is AI; the physics is exact. This architecture generates environments in which one experiences physics, not explanations about it. The child changes the lever arm, increases the load, sees the forces respond in real time.

The six-agent ensemble

The Mentor is an ensemble of six specialised agents operating on different time horizons. A Content-Agent generates in seconds. A Didactics-Agent adjusts difficulty over minutes to hours. A Navigation-Agent tracks progress over weeks to months. A Perception-Agent reads behavioural signals in real time. A Protection-Agent acts as a circuit breaker when frustration crosses from productive to destructive. An Audit-Agent monitors the Mentor itself for manipulative rhetoric, discriminatory bias or ideological drift.

Not everything in this ensemble is a large language model. Perception and Protection are fast classifiers; Navigation can be classical graph algorithms. The expensive foundation model is reserved for the expensive task: generating the nuanced, contextual learning experience.

Coordination follows a priority hierarchy — Protection overrides Didactics, which overrides Content. The Audit-Agent is architecturally independent: it runs in a separate execution context, cannot be overridden or disabled by the system operator, and its logs are append-only and tamper-evident.

One path produces citizens who own their learning. The other produces users who rent it.

How it begins

The Mentor works from the first interaction. No onboarding gate, no questionnaire. The learner says "I want to learn to read" and it starts — broad at first, narrowing with every interaction. Like any new relationship.

What the system cannot guarantee — socialisation, identity formation, the emotional weight of being seen by another human — it does not pretend to provide. Where teachers exist, the Mentor enriches their work; where they do not, learners still receive a complete academic education. The architecture replaces the instructional and assessment functions of a teacher. It does not replace the human being.

Reference

Architecture paper, Section 5, L1. DOI: 10.5281/zenodo.18759134. CC BY 4.0.