Calibra

Grades essays the way a specific teacher would — by comparison, learning their standard from a handful of anchor essays.
Demo judge · runs in your browser
A self-contained demo built to mirror Edexia's approach: comparative judgement + few-shot per-teacher calibration, scored with QWK. The grading maths (1-D Bradley–Terry placement, QWK) runs entirely client-side — no server, no API key.
Assessment task

Calibrate to:

Grade an essay

Try a held-out essay:

Accuracy on held-out essays

How closely Calibra matches Teacher A's real marks as it gets more of their graded examples. Metric: quadratic weighted kappa (QWK).

Switch teacher above and the curve changes — the same essays, calibrated to a different marker's standard. That's the "learns each teacher" idea, measured.