Theory becomes experience.

Germany is short 10,900 driving instructors. Average time to fill a vacancy: 307 days. The Deutscher Fahrlehrer-Verband reports that 34% of driving schools lose revenue directly because of the staffing gap. A class-B licence now costs €3,400 on average, climbing 5.8% per year (Destatis 2023).

The 45% theory fail rate sits downstream of that structural pressure. Students arrive under-prepared at their practical lessons, because classroom instruction — however thorough — cannot replicate the cognitive load of a real intersection.

The existing product landscape offers two inadequate answers. Cockpit simulators (€15,000–€40,000) require dedicated rooms and power lines that do not pencil out for a 5-to-15-person driving school. Passive 360-degree video asks nothing of the student. They watch; they don't decide. Neither addresses what BASt M 348 (2023/24) explicitly calls "concrete optimisation potential": the gap between theory instruction and the experiential test of hazard perception.

The regulatory frame sharpens the window. The FeV reform (expected H2 2027, supported by the VMK resolution of October 2025) will permit simulator hours within certified training, gated by a TÜV/DEKRA conformity declaration. Whoever is already piloting with Leipzig driving schools in 2026 sits in a different position than whoever starts after the reform.

One design decision was made early and held without flinching: Phase 1 requires no regulatory approval. By positioning the product as supplementary teaching material (not as creditable training hours), the team ships in 2026 without waiting for the FeV reform. This isn't a workaround; it's the correct sequence. Pilot data from Phase 1 becomes the evidence base for the certification filings in Phase 2.

The interaction model is neither quiz nor game. The student stands stationary inside a traffic scene on a Quest 3S. AI vehicles move around their position. The GazeDetector in Godot 4 with XR Tools validates the observation movements (shoulder check, intersection scan, sign recognition) before a decision question appears under time pressure. A three-star scoring system (0–3) separates observation score from knowledge score.

The marketing site runs on Next.js 16.1.6 with Firebase 12.9.0 as the backend for instructor-tablet monitoring. Three.js / React Three Fiber drives the ambient 3D hero. Framer Motion 12 carries the page transitions. The hardware target is Quest 3 / Quest 3S. Quest 2 is not supported. XR2 Gen 2 is the floor for the visual fidelity safety training requires.

Multi-agent workflow: Claude Code for architecture and business documents, Gemini for frontend UI, Codex for Firebase Functions. Each agent works in isolated branches with an explicit prompt contract before any merge. A shared .task-board.md tracks active work. The model permits parallel sprints without merge collisions.