Rocket Agent Lab

Track: Education

Why I built it

Rocket engineering is usually explained after someone else has already chosen the vehicle, payload and mission.

I wanted a lab where the learner could make the decision, break the design and see why it failed.

What I built

Rocket Agent Lab is an interactive 3D rocket classroom with inspectable vehicles, engines, structures, equations and launch tests.

Learners can change payload, engine health, wind, guidance and mission conditions. The simulator shows how those choices affect thrust-to-weight, mass ratio, delta-v, Max-Q risk and mission margin.

The clearest lesson is an engine-out experiment. The rocket configuration stays fixed while engine health falls from 100% to 67%. On the classroom Saturn V model, thrust-to-weight drops from 1.23 to 0.82, so the vehicle cannot lift off. Max-Q risk also falls, but only because the rocket never becomes fast enough. A smaller stress number does not mean the mission succeeded.

The tutor follows the evidence

The simulator creates a before-and-after receipt first. Only then does GPT Realtime receive the selected rocket, lesson, changed values and visible result.

An early voice version could respond with stale data when the Realtime channel became ready before the current scene state was attached. I rebuilt the context flow so opening and follow-up turns use the same bounded contract, with a fresh snapshot on every later question.

Tests reject stale thrust-to-weight, altitude and Max-Q values.

The simulator owns the numbers and pass/fail result. GPT-5.6 explains the consequence and suggests one next experiment.

How I used Codex

Codex helped inspect the browser wiring, isolate the Realtime context into testable modules, find the opening-turn race and build the deterministic proof card.

The main Codex build session is 019f5e81-8bde-79d1-b2d6-416776726a63.

What judges can try

Run the local app, open Flight and press Run engine-out. Inspect the receipt, then press Ask Realtime from this proof.

The rocket lab and proof card work without an API key. Live voice requires a server-side OpenAI key that is never exposed to the browser.

Current status

The deterministic causal lesson and no-network Realtime payload tests pass. The remaining gates are the authorized human microphone test and the public demo video.

This is educational software, not certified aerospace analysis.

Built With

• codex
• javascript
• node.js
• openai-gpt-5.6
• openai-realtime-api
• three.js
• webrtc