Fable Flight AI Instructor

Track: Education

Why I built it

Most flight simulators tell you whether you landed. I wanted one that can teach you why the approach became unstable, point to the physical control involved, and change the next lesson from evidence rather than generic encouragement.

Fable Flight is a browser flight school built around a real interactive simulator, a physical 3D cockpit, measured telemetry, and GPT Realtime instruction.

What works now

A public browser simulator with ten island regions, authored routes, weather, traffic, missions, replay, and multiple camera views. Blender-authored physical cockpit assets for the instrument pack, yoke, compass, engine quadrant, systems console, and flight controls. Eleven inspectable cockpit controls with measured values, normal bands, modeled failure boundaries, and small guided lessons. Voice-driven cold-start, pitot-icing, and alternator-recovery drills that require the learner to operate the real control; the instructor cannot secretly perform it. Black Box telemetry records high-signal events without storing microphone audio. GPT Realtime receives the same read-only aircraft, lesson, warning, and selected-control state visible to the learner. Deterministic tests reject impossible or ungrounded coaching states.

A representative release check on July 18 first exposed a one-to-three-pixel layout mismatch. After accepting a three-pixel rendering tolerance, the rerun passed all nine gates: mission selection, measured storm inputs, concise operational strip, balanced desktop geometry, cockpit continuity, Realtime grounding, Black Box handoff, non-regulatory wording, and zero console errors. Human review still found the physical cockpit composition too dense, so passing geometry tests are not presented as finished visual design.

Why Codex and GPT-5.6 matter

Codex has been the engineering collaborator across telemetry contracts, cockpit assets, browser verification, accessibility, map behavior, and Realtime grounding. GPT-5.6's Build Week role is the adaptive instructor: identify a repeated pattern across measured events, explain it in the learner's current situation, and choose the next drill. Deterministic code remains responsible for aircraft state, safety boundaries, and whether an event actually occurred.

Build Week implementation

Every capability presented in this submission was built after July 13 during Build Week: the browser flight school, physical cockpit, evidence-linked coaching, physical-control lessons, telemetry-grounded Realtime tools, debrief contracts, and verification harnesses. Unrelated repository shells, references, or experiments are not part of the judged claim.

The demo I need to prove

One learner flies an unstable circuit. The debrief cites two exact events. The learner asks why. The second circuit changes one lesson based on that evidence, and the verifier proves the instructor cannot cite an event that never happened.

Current judge path

The public simulator is available now. The current judged implementation is anchored in commit a791496, with BUILD_WEEK_SCOPE.md recording the post-July-13 submission boundary. Before submission I still need to simplify the dense cockpit overlays, record the two-circuit human Realtime test, add the exact /feedback Session ID, finish the README/provenance checklist, and record the public under-three-minute demo.

Human-edit note

This is a working draft grounded in the current build. I will rewrite the final description in my own voice after the human teaching test.

Built With

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