Defy the Vector: what a frontier code model builds when the brief is a planet.

The brief handed to the model was one page: build body-free first-person flight that starts at eye level in a Washington garden, crosses the atmosphere in one continuous move, and arrives at a navigable Earth, on real geography, with no avatar and no loading wall. City demos rarely leave the city and globe demos rarely start at human scale, so the seam between the two scales is exactly where generated code tends to fall apart. That is why it was chosen.

Two constraints made the brief measurable rather than impressionistic. Every build had to pass an automated budget gate, and every claim in the eventual report had to trace to something in the recorded sessions. The experience is the instrument. The model is the subject.

The first prompt was written on a Friday evening in July 2026. A budget-gated production build existed before morning. The sessions ran as supervised direction: bounded prompts, one decision at a time, and a human engineering review gate on every output before it entered the tree. Nothing merged on trust.

Inside that loop, GPT 5.6 Sol carried the implementation. It proposed the state machine that separates the ground field, the ascent, and the orbital field, wrote the movement model with inertia and altitude-scaled speed, and kept a single coherent architecture across the whole window instead of drifting into patchwork.

• Supervised sessions with bounded prompts and recorded decisions
• A human review gate on every output before it entered the build
• An automated budget gate failing any build that overran its limits

The measured output is roughly 3,500 lines of TypeScript across eighteen modules, about 5,600 lines of source overall once the page shell, styles, and build tooling are counted. The architecture is genuinely good: the planet, its imagery, and country boundaries load only when a viewer chooses to ascend, so a visitor who stays in the Washington field never downloads the Earth. That single decision, which emerged early in the sessions, is why the first payload sits at 179 KiB gzip against a 240 KiB budget.

The mathematics held up under re-testing. Orbital travel interpolates along the shortest spherical arc between real coordinates, distances derive from angular separation, and re-entry returns the viewer to the same Washington field they left. Speed and volume are not the story. Structure under constraint is.

Four material corrections were logged in the window, and none of them were syntax. The model shipped relative asset paths that failed once the experiment was deployed under a nested route it had not been told about. It could not judge which Earth textures carried confirmed redistribution rights, a call that decides what may legally enter a public repository. Its first attribution layer was silently stripped by the minifier chain and had to be rebuilt in depth. And one deployed copy went out without its texture set until inspection caught the gap.

Every one of those corrections is a judgment call about context the model did not hold: deployment topology, intellectual property, provenance, and verification. That boundary, not code volume, is what this experiment was designed to locate.

The Washington environment is geographically grounded but deliberately lightweight, built from bounded civic geometry with a few procedural landmark forms for orientation. It is not a survey, a navigation aid, or a claim about any specific structure. Treating a legible model as an authoritative one is exactly the mistake this program exists to avoid, so the interface and this report both name the boundary.