An inspectable AI reasoning workspace that protects student authorship: memory, lenses, limits, counterviews, and causal replay before the learner writes.
Inspiration Generative AI can make a student answer look polished while hiding how the student got there. PRISM starts from a different premise: personal memory can shape a question, but it is not evidence or universal truth. ## What it does PRISM is an inspectable reasoning workspace for education. A learner adds a Memory Card, sees it explicitly labeled MEMORY · NOT FACT, and then explores a question through lenses, declared break points, counter-lenses, and open evidence questions. Remove-and-Replay recomputes the reasoning packet after removing a memory so the learner can inspect causal influence. The final position remains an empty author field: PRISM does not write the essay. ## How we built it We used Codex to structure typed contracts, implement the reasoning and replay flow, test the product, and iterate on the demo. GPT-5.6 is used in the evaluation/reasoning workflow to produce structured, inspectable analysis signals. The live demo defaults to clearly labeled Fixture · $0 mode; the evaluation panel separately labels synthetic GPT-5.6 evidence and its limits. ## Challenges The hard part was making influence visible without turning private meaning into fact or turning the model into the author. We solved that with explicit provenance labels, break points, counterviews, replay, and a refusal to generate the final submission. ## Accomplishments A working browser demo shows the full Memory -> Inspect -> Illuminate -> Author flow and causal replay. The public repository includes the implementation, a MIT license, and a clear testing path. ## What we learned A credible education AI demo needs to show both capability and restraint. The strongest proof is not a fluent answer; it is a learner being able to see what shaped an interpretation and decide what they believe.
Log in or sign up for Devpost to join the conversation.