Every surface, in the order it fires.

The Socratic tutor. Attempt first, guided second, answered never.

Every academic conversation on Brainback runs inside Learn mode. The system prompt is versioned, cached, and enforced server-side. It refuses to output final answers. It refuses to work the problem for you. It hands back a question, one correction, and the next hint you've earned.

Five calibrated rungs. None to the answer. The final step is always yours.

Concept, then tool, then setup, then first step, then second-to-last. That's the ladder. Level 6 doesn't exist. Every hint costs points. A hint that costs nothing becomes an answer button — which is exactly what we refuse to be.

Teach it back in your own words. Haiku scores 0–5 with actionable feedback.

Getting the answer is the easy half. The hard half is being able to reconstruct the reasoning from scratch a week later. Explain-back forces the reconstruction on the day it counts most — right after you got the concept — so the memory encodes deep.

SM-2 scheduling. Cards generated automatically from every wrong attempt.

You never write your own flashcards. Every misstep in a Learn session becomes a card, tied to the concept it revealed. The queue expands and contracts by the SM-2 algorithm — one of the most-replicated findings in learning science, made operable.

One weekly number. Recall accuracy, confidence calibration, explain-back quality.

40% recall accuracy, 30% calibration, 30% explain-back. The score is a compass, not a leaderboard. It moves slowly. It resists gaming. It tells you whether the semester is compounding — or leaking.

Normal AI for everything not in your declared subjects. Full streaming, full completion.

Rewriting your résumé, brainstorming a talk, coding a scratch script for a project outside your major — Efficiency mode handles those the way any other assistant does. Learn mode never gets crossed with it, and the tutor never leaks a solution through the efficiency lane.

Every reply passes a second-model check before it reaches you. No leaks. On purpose.

A single system prompt is easy to jailbreak; a separate second-model output guard is not. Every tutor reply is re-read by Claude Haiku with one job — was this a covert answer? — and regenerated or refused if it was. The guarantee is architectural, not textual.