How this was built
Caelus was written almost entirely by AI coding agents. The Python reference, the TypeScript port, the tests, the website, and most of this prose came from agents working under human direction. That is only worth saying because of the second half: every number the engine produces is checked against two independent references, and a golden suite fails the build the moment the code drifts. The checking is the reason an engine written by machines can be trusted.
Reference-first, checked to the last digit
Every model landed in the Python reference first. That reference writes golden fixtures: positions and events at fixed instants, stored as JSON. The TypeScript engine is a port of the same math, and a conformance suite replays the fixtures against it on every commit, to the last digit the fixture records. A model is not shipped until the reference, its fixtures, and the port agree. An agent can move quickly when a red suite catches the exact moment a port stops matching its reference.
Two oracles, not the author's word
Positions are checked, not asserted. Swiss Ephemeris 2.10 is the same-frame oracle across 1900–2099; JPL Horizons is the independent reference at a separate set of epochs. The per-body bounds are published on Validation, and the methodology behind them on Methods. For an engine written by a machine, the trust comes from the oracle rather than the author, which is the same standard a human-written engine should meet.
Direction stayed human
The agents wrote the code; a person chose what shipped. Work lands on a development branch, and a person promotes it to the main branch, so someone signs off before anything reaches users. Dead-ends were reverted rather than shipped: when a heavier lunar data pack broke the browser bundle and lost accuracy against the reference, it went back to the existing fit, which already sat at the reference floor. The agents supplied speed; the judgment about what was correct and what belonged stayed with a person.
The prose has the same gate
AI writes like AI. The repository runs a prose linter with a rule set for the usual machine tells: the em-dash used as a connective, the three-verb drumroll, the manufactured conclusion, the closing pleasantry. It fails continuous integration the way a failing test does. This page was held to the same gate.