TeachBack

The description states that TeachBack is a tool for educators to scaffold student learning by requiring students to explain concepts to a simulated peer named Nova, who holds a plausible misconception. The system parses source documents and constructs:

• A concept map (3–5 nodes)
• Literal evidence anchors from the source
• One target misconception for Nova
• Student prompts

The workflow includes:

1. Instructor submits a source document and learning objective.
2. System builds baseline setup, which instructor approves.
3. Student explains to Nova and applies reasoning to a novel scenario.
4. Nova’s internal state updates only when student addresses causal links with evidence.
5. Entire trace is submitted for instructor review.

The system uses two structured generation phases:

• Setup Phase: Concept map, evidence anchoring, misconception formulation.
• Turn-Update Phase: Nova replies, validates evidence grounding, modifies concept-map state.

It avoids automated grading or AI-detection mechanisms and enforces strict server-side validation to prevent hallucination or model drift. The architecture is described as deterministic for offline demonstration purposes.

Evidence Self-reported by the authors; no external verification provided.

The description states that TeachBack reframes the challenge of AI in education from a policing problem (e.g., detecting cheating) to a constructive, formative practice exercise. It positions itself as an agentic clinical coach for learners, emphasizing peer-tutoring effects and cognitive scaffolding.

Key claims:

• The tool avoids algorithmic surveillance or AI-detection arms races.
• It explicitly avoids automated grading or epistemic claims about internal knowledge.
• It focuses on making the AI’s mechanics legible through visible source maps, real-time state changes, and declarations of uncertainty.
• The system is built to maintain instructor authority via mandatory pre-approval and post-activity review.

Evidence Self-reported; no third-party validation or market positioning data provided.

The description identifies instructors as the primary users. Instructors submit source documents and learning objectives, approve baseline setups, and review student traces.

It also implies students are the end-users of the system, though not explicitly defined in terms of demographics or grade levels.

Evidence Self-reported; no explicit segmentation or targeting data beyond “instructor” and “student.”

Not evidenced. The description does not mention any pricing model, monetization strategy, or business model.

The system is built using:

• Framework: Next.js
• Language: TypeScript
• AI Integration: OpenAI Responses API with GPT-5.6
• Architecture:
  • Two structured phases: Setup and Turn-Update
  • Strict server-side validation to prevent hallucination or unauthorized state changes
  • JSON schemas for grounding generative output
  • Deterministic offline demo architecture

The codebase includes automated test suites covering workflow gates, citation validity, malformed inputs, and state-update boundaries.

Evidence Self-reported; no information on deployment, scalability, or infrastructure beyond the development stack.

Not evidenced. There is no mention of:

• Revenue
• Customers
• Users
• Adoption metrics
• Product usage data
• Market traction

The project is described as a hackathon submission and prototype, with no indication of prior use or iteration in real-world settings.

Not evidenced. The description does not reference existing tools, platforms, or competitors in the educational AI space.

1. Unvalidated Pedagogy: The approach is described as a novel pedagogical loop but lacks empirical validation or classroom testing.
2. Limited Scope: The system avoids features like rubrics, LMS integrations, and student accounts—this may limit its utility in real-world adoption.
3. Self-Reported Only: All claims are self-reported; no third-party verification, user feedback, or performance metrics are provided.
4. No Commercialization Path: No indication of how the tool would scale beyond a hackathon prototype or how it would generate value for paying customers.

Inference The lack of real-world testing and commercial viability raises concerns about whether this is a viable product or just an experimental idea.

1. Has the system been tested with actual instructors and students in real classroom settings?
2. What specific pedagogical outcomes have you observed from using this approach?
3. How do you plan to validate that the misconception design and prompt structure improve learning?
4. Are there any plans for integrating with LMS platforms or scaling beyond the prototype?
5. What is your roadmap for moving from a hackathon project to a product that could be used in schools or institutions?