AI tutors give in and hand out answers when students push hard enough.
A Socratic tutor in 35 languages that guides with questions and holds under pressure.
Real compliance raised from 22.2% (despite a 94% evaluation score) to 100%.
The problem
An AI tutor may be designed to teach exclusively through guided inquiry, yet under pressure — student frustration, parental demand, deliberate circumvention attempts — many such systems default to direct answers, undermining the pedagogical model precisely when it is most required to hold.
Who it is for
- Students and families requiring conceptual comprehension, rather than task completion alone
- Educational institutions and platforms requiring a tutoring system that maintains its pedagogical methodology under authentic adversarial pressure, not merely within controlled demonstration conditions
Market context
Khanmigo (Khan Academy) constitutes the established, dominant benchmark for Socratic AI tutoring — deployed across 700,000+ K-12 students and 380+ district partnerships. Additional established competitors include ChatGPT Study Mode and Google's Socratic (which, despite its designation, predominantly delivers direct answers rather than Socratic inquiry), alongside subject-specific tools such as Photomath.
The gap it closes
Independent analysis of the category's own benchmark identifies material limitations directly: "Khanmigo is weak at math (LLM limitation) and weak at keeping students engaged." More broadly, the category's defensible competitive advantage is explicitly identified as "the Socratic constraint, the safety monitoring... not the model" — establishing that architectural discipline in maintaining the teaching method matters more than the underlying model, and that maintaining this constraint under pressure remains the category's principal unsolved problem.
How it works
- A student submits a query in any of 35 supported languages
- ARIA responds exclusively through guiding questions — direct answers are architecturally excluded, without exception pathway regardless of framing or pressure
- Compliance is continuously validated against an adversarial golden dataset, including frustration-pressure and circumvention-pattern test cases specifically engineered to breach the constraint
What makes it different
- Zero exception paths, by deliberate architectural constraint — validation demonstrated that even a narrowly-scoped exception becomes the default circumvention route under sustained pressure
- Adversarially verified, not merely asserted — a documented 94% automated evaluation score against 22.2% real-world compliance was identified and resolved to 100% through rigorous testing
- 35-language deployment — extending access beyond single-market, single-curriculum constraints
Architecture
System Overview
Component Breakdown
| Response Generator | Structurally incapable of emitting a direct answer, not filtered after |
| Deterministic Pattern Layer | Fast, free first-pass check for obvious violations |
| deepeval GEval Judge | Catches subtler, paraphrased answer-leaks |
| Manual Review Queue | Unsettled calls routed to a human |
Data Flow
- A student submits a question in any of 35 languages
- The generator produces only guiding questions — no code path for a direct answer exists
- Every response passes 3 evaluation layers, ending in human review for unsettled cases
- Golden dataset testing continuously re-validates compliance
Technology Stack
| Vector storage | ChromaDB | Retrieval grounding for curriculum-aligned responses |
| Compliance evaluation | Deterministic + deepeval GEval | Two different detection methods, different failure types |
| Language support | 35 languages | Multilingual jailbreak resistance explicitly tested |
Real Design Trade-off, Documented
An exception for 'simple factual questions' was considered and rejected — testing showed it becomes the path of least resistance under pressure, since almost any question can be reframed as simple.
Performance, Measured Not Estimated
- A real gap closed: 94% automated eval vs. 22.2% real-world compliance, closed to 100%
- 1 real regression caught on first full run, by the golden dataset, not a user report
See ARIA working
A guided walkthrough, one-on-one. We usually reply within one business day.