Short films about Oracle7. Every number on screen is a measured held-out result. Anything simulated or not yet built is labelled on screen.
The spiral from 6.0 to 20.0, filling in as each version passed its sealed test, with the failure left on it; the build roadmap with its estimates; the vision of the Slow Fire; and what the dates do and do not mean.
A live session on the released 10.1: Claude hands it 26 repeated tasks, then 19 of its own answers written from memory before anything ran. Recorded in a terminal; long waits shortened and marked.
Two jobs beside another agent, what it does on its own, and the advantage: answers that can be checked again, an error rate that is measured, and failures that differ from a language model's.
Every internal version since 4.9 joined in one release and tested again as a whole.
The released tree in one session: it verifies itself, talks, writes an essay, and makes a film for a problem. The real output.
From one laptop to every connected machine: inside the model, the family, deep space, a city routed by trafficcop, sensors on farms, factory lines, watchtowers and buoys, and the one you talk to.
Download MP4oracle7-deepspace, flown against a simplified safe-mode baseline on seven real spacecraft at their JPL Horizons positions.
One real failure, “ok thanks”, followed from the failure inbox through the fold worker and the exam to fold w472.
Six models, each with its measured card, and the numbers that are still red left on.
Twenty-four hours of a self-improving model, as a clock around the Earth: failures recorded, the 03:30 benchmark, repairs through the gate, and every service following the chain.