Epicycles.
When the corrections you have to make to a model start to multiply, the fix is not a better correction. The fix is to reconsider the ground truth you assumed. Ptolemy did not need more epicycles; he needed Copernicus.
The puzzle
[DRAFT — your claim] For fourteen centuries, Ptolemaic astronomy explained the motion of the planets by stacking circles on circles. The base orbit was wrong, but you could patch it: add an epicycle, then another, and the predictions kept matching the sky. The fit was excellent. The story was a mess.
The fix
[DRAFT] Copernicus did not propose a better epicycle. He proposed a different ground truth: the Sun, not the Earth, at the centre. The new picture explained retrograde motion not as a real wobble but as an artifact of the observer's frame. The fit was not initially better. The story collapsed into simplicity.
The flip
[DRAFT] The lesson is not "models are bad." Ptolemaic astronomy was a working predictive model for over a thousand years. The lesson is about what to do when the corrections start to multiply. Every additional epicycle you add to keep the predictions in line is a signal — not about the data, but about your frame.
The point
[DRAFT] Definitional precision matters because it tells you which kind of fix you need. If you do not know whether you are debating a model (the operator) or its ground truth (the frame the operator is fit inside), every disagreement turns into a fight about epicycles. Lectures 01 and 02 are the language for telling those two things apart.
Epicycles in modern model risk
[DRAFT] Most "AI governance" today is epicycles. The base assumption — that a model is a function — is wrong (Lecture 01). The fixes pile on: bias dashboards, drift alarms, fairness audits, red-team protocols, alignment patches. Each one is locally reasonable. The accumulation is Ptolemaic. The Copernican move is to redefine what is being governed: not the function, but the operator-in-context, and the trust between it and its observer.