Real, deterministic assignment generator — no LLM judgment anywhere in the grading. A problem is generated, a numeric answer is computed by code, and your answer is compared to it directly. Data is stored only in your own browser (localStorage) — nothing is sent anywhere.

Calibration exploration bug fixed, stated directly: a real forward-walk test previously found that the difficulty-selection fallback was a fixed function of the target alone — it never sampled a second tier once one was in use, so it couldn't discover a better fit for students far from average ability. This has been replaced with a real Wilson-score-interval gate: a tier only counts as confirmed once it has at least 5 real attempts and its confidence interval genuinely contains the target rate; below that, the system actively explores the least-sampled tier instead of repeating the same one. Tested directly against 5 real scenarios — including a check that a single lucky attempt can't prematurely lock the system onto the wrong tier — before this replaced the old mechanism. Full detail, including the real capability this reuses: Capability Library · Backtest-Only Audit.