The visitor's camera
A runbook from someone else, our own numbers, and a change of direction
Someone else built a game with the same tools in the same fortnight and wrote down what it took: 39 tracked hours, 1.56 billion tokens, 90 commits, a hero character with a face rigged from a generated mesh, and the admission that 80 percent of the work was polish. I read it twice and then measured us the same way.
Our two seats had used 444 million tokens across 36 sessions in four days, and we had never counted. Our records are stricter than his in every other way: 70 decisions, eight reports, 7,097 saved renders, 285 paid generation jobs each with its input hashes. But every one of those renders was judged from four aerial cameras. His was judged from where the player stands.
So the direction refined rather than restarted. Two eye-level cameras join the four: E on the beach, F in the middle lane of the fishing town. Finding F took eight renders, because a camera at two metres in a palm grove sees fronds first. The world will start to move on its own clock: fronds, smoke, birds, surf, cattle. And one fisherman will go through a character route we already have, a three-view sheet, a mesh, a remesh, a rig, a walk.
Phase 09 opened with those foundations attached to the last aerial fix. This leaf is being written while it runs.
- learnings, the study of another builder's runbook, 13 September
- D-2026-09-13-02, Stage 1b proposed
- D-2026-09-13-03, phase 09 re-briefed

