Adaptive Learning Companion

The premise

Two findings in learning science have survived decades of replication: testing yourself beats re-reading, and spacing practice beats cramming. Almost no consumer learning app is honest about the third finding — that effective practice feels worse than ineffective practice. Streaks and confetti optimize for the feeling of progress. This concept optimizes for retention and accepts the product consequence: sessions that are short, effortful, and occasionally humbling.

How adaptation works

The scheduling engine is classical and inspectable: a spaced-repetition model tracks per-item memory strength from response accuracy and latency, and difficulty moves within a band — struggling learners get easier retrieval cues and smaller steps, strong performance pulls items forward and interleaves related topics. No neural network decides what a learner sees next; the adaptive core is auditable arithmetic, and a learner (or parent, or teacher) can always ask the honest question “why this card now?” and get a real answer.

What the AI tutor may do

Language models enter in one role: explanation on demand. When a learner misses an item, the tutor can rephrase the concept, generate a worked example, or answer “but why?” — grounded in the course’s own vetted material, not the open internet.

Hallucination is handled structurally, not aspirationally:

• Explanations are generated against retrieved passages from licensed or author-provided content; the source is shown with every explanation
• If retrieval finds nothing relevant, the tutor says so and links a human-written reference instead of improvising
• Generated practice variants are drawn from templated transformations of vetted items, not free composition — novelty in surface form, not in facts
• A standing red-team set of known-tempting errors per subject runs against every model or prompt change

What the AI tutor may never do

It never grades high-stakes work. Essays, exam preparation submissions, anything that feeds a real academic consequence — out of scope, permanently. A model’s confident misjudgment of a twelve-year-old’s essay does damage that no accuracy percentage excuses. The tutor also never assesses ability (“you’re a visual learner”, “you may have difficulties with…”): it schedules items and explains content, and the vocabulary of diagnosis is banned from its output by policy and by test suite.

Parent, learner, and the data between them

For minors, a parent sees process, not surveillance: minutes practiced, topics covered, schedule adherence. Item-level struggles — which questions were missed, what the learner asked the tutor — stay with the learner. A child who knows every wrong answer is reported home stops practicing where it’s needed most; the product protects the psychological safety that makes retrieval practice work. Adults own their data entirely; tutor conversations are not training data; deletion means deletion.

System overview

• Content pipeline: authored or licensed item banks with per-item provenance; a review workflow gates anything AI-assisted before it reaches learners
• Scheduling engine: server-authoritative, offline-tolerant — practice works on a plane, syncs later
• Tutor service: retrieval-constrained generation with per-response source attribution and the red-team suite in CI

Prototype scope

One subject with objective right answers — vocabulary or basic mathematics — one learner age band, the scheduling engine, and the retrieval-grounded tutor. Subjects where correctness is contestable (writing, history interpretation) wait until the grounding discipline is proven where it can be checked.

Evaluation plan

• Retention, measured properly: delayed testing at two and six weeks against a self-study control, on matched material
• Tutor grounding audit: zero unsourced factual claims tolerated across a 500- explanation sample; error rate on the red-team set tracked per release
• Effort honesty: session completion and return rate for a product that refuses confetti — the commercial hypothesis under test
• Boundary integrity: interviews with parent–learner pairs on whether the visibility split builds trust on both sides or satisfies neither

Open questions

Whether evidence-based practice can retain consumers against products optimized for feeling productive is the existential question — it decides marketing, pricing, and whether the concept should be sold to families at all, or to schools where the incentive to measure real retention already exists.