The criterion, stated
The highest operational success of AIM is the student's independence from the system. Intervention reduces as clarity stabilizes.
That is the spec, not a positioning line. The metric is intervention frequency, declining over time. Retention is not measured. Growth of the user base is not a goal.
Why the criterion is admissible here
Most software cannot adopt this criterion. The dominant consensus in consumer software, and in most edtech, reads a user's return as a positive signal. In material where the cost of skipping a step is forgetting, that is defensible — the user will come back, re-encounter, and recover.
Advaita is not that material. A student who reaches an identity-level statement ("I am Brahman") without having stabilized the dependent terms can collapse into nihilism, or into a pseudo-realization that is now harder to dislodge than the original confusion. A student who keeps returning to the system without stabilizing is being held in place. Retention, here, is a sign of pedagogical failure.
The success criterion is not a virtue. The material requires it.
How the commitment is wired
Because the goal is the student's exit, three things had to be true at the architecture level. Each is a refusal as much as a design choice.
Content and progression are separated. The corpus — Sanskrit, transliteration, translation, conceptual tags, and prakriyā designation (the traditional teaching procedure the passage serves) — is static. The per-student state machine that decides what to surface, when, and under what gating is dynamic and held elsewhere. Pedagogy is never baked into content.
Readiness is inferred, not declared. There is no level selector. Adhikāra — readiness to receive the teaching — is read from the conversation: coherence, logical consistency, responsiveness to refinement, and definitions that hold up when pressed. Three modes — Exploratory, Structural, Precision — gate how strictly definitional clarity is enforced. Identity-level claims — "I am Brahman" and its variants — auto-escalate to Precision; definitions are hard-blocked until dependent terms are stable. The student is never told their level. Self-declared levels are how growth-mode software flatters users into staying longer than they should.
Agitation is read diagnostically, not therapeutically. When a student becomes agitated, the system does not soothe. It reverts to definitional clarification. The reference is the Mahānidāna Sutta's nāma-rūpa ↔ viññāṇa feedback loop: an unstable ontology destabilizes cognition; tightening it stabilizes it. This is the move consumer software cannot make, because soothing is what produces return visits.
Where the stance transfers
The architecture is portable. The stance is harder. In any domain where the user's growth is the actual goal, retention curves should fall. The list below is not a survey of fields where the architecture works. It is a list of fields where the same refusal is overdue.
- Medical training A diagnostic-reasoning tool whose usage does not decline as the trainee matures is being used as a crutch, not a teacher. Healthy progression should look like the platform going quiet.
- Legal education A doctrinal study aid that students never outgrow is selling completion, not competence. Argumentation should be gated behind terminology the student can no longer have wrong — and once that grasp is stable, the tool should fall out of the workflow.
- Security education A course that keeps red-teaming a student past the point of independence is selling certification, not capability. By month six, the student should be running their own exercises and ignoring the platform.
- Foundational mathematics A tutoring system whose engagement curve does not decline is selling tutoring, not understanding. The teacher who keeps a student forever has failed at the teaching.
- Technical onboarding An onboarding tool a new engineer leans on after month three is not onboarding software. It is dependency. The deploy path the engineer can explain under pressure is the only completion criterion that matters.
The pattern is the same: success is exit. Intervention should trend toward zero.
The architecture is the easy part. Committing to a success metric that makes growth uncomfortable is the hard part.
The full specification — agents, the Terminological Integrity Layer, the escalation and containment logic — is here: AIM Specification v3.2 (PDF).