What should you learn next?

The actual Core Kanji scheduler (the one proven in the simulation spike) running live on a toy corpus. Pick a subject and drive it.

Most study tools march you through a fixed order and, after each card, ask you to grade yourself. This runs the real engine instead. It:

Tracks what you know as real FSRS memory state (stability, retrievability, a review schedule), not a checkbox.
Picks what to teach next by corpus frequency × word-completion, building each new combo on ones already mature.
Grades by latency, never self-report. Correct-and-fast, correct-and-slow, and wrong become Easy / Good / Hard / Again, inferred from your answer time against your own rolling baseline.
Is still spaced repetition. The FSRS clock schedules every review underneath; this decides what to surface and grades it for you.

Same engine.js under all three. Its numeric core is parity-tested against simulate.py. Toy corpus; the mechanism is real.

A demo of the Core Kanji learning engine (AGPL-3.0-or-later). Source, engine, and this page: github.com/moxzas/core-kanji (/demo). engine.js is the real scheduler; demo.js is just its UI.