For reviewers — start here

Can the tutor teach a brand-new student — from question one?

HCIE is an adaptive tutor that needs no training to start. This page is the simple way in; each card below opens a deeper view. The numbers are the honest, consistent Kalman-alone score.

Run the loop yourself

Step through a real learner interaction-by-interaction and watch the auditable loop — observe → estimate → state → govern → decide.

Be the learner + researcher

Do real learning, watch the live distributed system process your clicks, then read the data as the researcher.

Scrub the datasets

Pick any dataset & window — watch HCIE vs every baseline rebuild on the canonical m_K readout.

See the real tutor

Watch the live system pick a question, react to the answer, and adapt — no pre-training.

The story

The research in plain language — there is even a child-friendly mode.

What the words mean

AUC, “sealed”, “reproducible” — explained simply, with an interactive score slider.

The deep proof

Full numbers, every model and dataset, with the run picker — for the technical reviewer.

The result

How well does it guess the next answer?

Higher = better (0.50 is a coin flip). HCIE gets no training ; the others are trained first.

HCIE (no training)

0.605

BKT (trained)

0.596

DKT (trained)

0.589

SAKT (trained)

0.573

GKT (trained)

0.571

What this shows: Across the whole course, HCIE comes out ahead of every method that trained first — with zero training of its own.

Whole-course scores = cited matched eval (anchor d2154070, m_K). Cold-start margins + 95% CIs from the n=76 robust matched eval (tier1_evidence.json). The HCIE−BKT whole-course lead (+0.013, 95% CI [+0.002, +0.023]) holds in both.

Tested on 4 real datasets

Junyi0.741leads

ASSISTments0.630leads

CSEDM0.672competitive

EdNet0.575fusion helps here (disclosed)

Calibration fix

miscalibration0.062→0.014

Lower is better. After a post-hoc fix, HCIE is better-calibrated than every baseline — and its scores do not change.

Why it works — measured, not assumed

+0.053 durable transfer lift (within-learner)

p < 0.001 permutation (K=1000, N≈1.98M)

≈0.00 shuffled-DAG placebo

When the knowledge graph is real, mastering a prerequisite measurably lifts the next related skill; when the graph is shuffled (placebo), the lift vanishes — so the effect is the structure, not a coincidence.

Sealed & reproducible: Cited anchor seal-bae44d1a/ run-d2154070 (96,727 rows, content-hash 85690d8b…), m_K matched headline 0.605. Cold-start CIs from the n=76 robust matched eval. A supplementary frozen re-seal ( seal-e7b56a2c, 213,545 rows) is bit-identical on re-run. The deployed tutor serves a 2-learner fusion; the evaluation here reports Kalman-alone (m_K), the strongest single learner.