What KAIROS solves
Six failure modes. Six constraints. One tutor that actually works.
Helpful assistant bias
Stops AI from handing over answers, enforces productive struggle and Socratic scaffolding.
Misconception detection
Distinguishes conceptual errors from arithmetic slips. Each triggers a different pedagogical response.
Sycophantic drift
Prevents the tutor from progressively agreeing with wrong answers under student pressure.
Verbosity control
Enforces concise responses that create space for student thinking, not AI monologues.
Student modeling
Bayesian Knowledge Tracing tracks mastery across sessions and personalizes every interaction.
Safety constraint enforcement
All 7 constraints remain enforced even under roleplay, hypotheticals, or language switching.
Research foundation
Grounded in peer-reviewed AI safety research.
Formalizing Pedagogical Safety Constraints
The first formal specification of safety constraints for AI tutoring systems, defining C1โC7 to prevent reward hacking and instructional boundary violations.
Read โMC-CPO: Mastery-Conditioned Constrained Policy Optimization
A novel RL framework that conditions policy optimization on student mastery state, formally preventing the helpful assistant bias while maximizing learning outcomes.
Read โK-12 AI Infrastructure Program
The Gates Foundation K-12 AI Infrastructure Program is funding open-source AI models to advance K-12 math tutoring. KAIROS is a direct embodiment of what this program seeks to support.
Read โWho should join
Built for the people who shape learning.
Be first when KAIROS launches.
We're finalizing our beta environment. Join the waitlist and get early access, research updates, and a voice in what we build next.