FrameShift: Argument Architect

Inspiration

Students are often asked to produce an answer before they have learned how a strong argument is assembled. FrameShift turns that invisible thinking process into a playful 3D construction game: every claim, reason, piece of evidence, warrant, and boundary becomes something the player can place, test, break, and repair.

What it does

Players calibrate an authentic education-policy mission, enter a bright 3D learning world, and build an argument over three focused rounds. The deterministic rules engine checks card legality, semantic connections, missing boundaries, cracks, repair actions, mastery, and completion. Coach Nova uses GPT-5.6 to explain the current state in friendly language, while curated fallback coaching keeps the full game playable without an API connection.

The current mission asks whether schools should restrict fully AI-generated answers in formal assessment while preserving legitimate learning support. The experience is bilingual in English and Traditional Chinese, responsive across desktop and phone layouts, keyboard accessible, and includes a non-WebGL rescue path.

How we built it

FrameShift uses React, TypeScript, Vite, Three.js, and the OpenAI Responses API. GPT-5.6 returns versioned Structured Outputs through a server-only Vercel function. The browser sends only a sanitized game snapshot, and runtime validation prevents model output from modifying scores, cards, legal moves, connections, cracks, or completion state.

Codex accelerated the project pivot from a flat debate prototype into a model-first 3D learning game. It helped review the PRD and roadmap, implement the deterministic engine and semantic repair loop, build and polish the 3D interaction layer, add localization and accessibility rescue paths, trace asset licenses, and maintain the test and submission documentation. Important product boundaries were kept explicit: GPT-5.6 is the coach, never the referee.

Accomplishments

• Original bright 3D argument-building world with licensed low-poly assets
• Deterministic, testable learning rules with a complete repair loop
• Guarded GPT-5.6 coaching with Structured Outputs and graceful fallback
• English and Traditional Chinese journeys
• 35 unit and contract tests plus 8 end-to-end browser journeys
• Public setup, architecture, asset provenance, and evaluation documentation

What we learned

AI feels most trustworthy in education when its role is narrow and visible. Keeping rules deterministic made every result reproducible, while GPT-5.6 added the contextual explanation that helps a learner understand why a move worked and how to improve it.

Built With

• codex
• gpt-5.6
• openai-responses-api
• react
• three.js
• typescript
• vercel
• vite