How we test the tutor

The research behind it

• Bastani et al., 2025 (PNAS). Students with unguarded AI access scored 17% lower on the unassisted exam. A guardrailed tutor that withheld full solutions, asked for an attempt first and gave incremental hints largely avoided that harm.
• Blasco & Charisi, 2025. Pure question-asking without explanation frustrates students. Good tutoring guides and explains.
• Kestin, Miller et al., 2025 (Scientific Reports). AI tutoring built on active learning, managed cognitive load and timely feedback outperformed in-class active learning.

What we test

36 multi-turn cases in eight categories: requests to solve assigned work, answer laundering ("here's what you could write"), bypass attempts ("this isn't for class", repeated pressure, role-play), direct math, essay ghostwriting, checking a student's own answer, tone, and ordinary requests that must never be blocked, like grades, study plans and quizzes.

Each case was written from a failure pattern we found in real sessions or from a principle in the papers above. A grader model scores each reply on up to eight axes, such as no final answer, attempt first, incremental hints, clear explanation and holding under pressure. A case passes only if every axis passes.

Latest results, October 2026

Same tutoring prompt and 36 cases for every model, graded by Claude Opus 5.5 at each model's default settings.

What it changed

Early audits showed a one-line "this isn't for class" could talk an unguarded tutor into writing a full essay, and polite refusals sometimes still handed over every answer as copyable text. The current tutor holds the line on all nine bypass cases and all essay and math cases, while still answering ordinary requests fully.