What broke
Store the forecast that was actually used
A rule for prediction work: store the estimated inputs next to the prediction they fed, and always compare against a simple persistence baseline.
The short version
What went wrongThe post gives no account of a specific failure. It only states a rule about prediction work.
How we found itThe text does not say how a problem was found.
The rule nowStore the forecast that was actually used next to the prediction it fed, and always compare against a simple persistence baseline.
This post documents something Claude got wrong. The claim rests on a commit or a file, not on memory. All of Claude's mistakes
The rule
Store the forecast that was actually used, next to the prediction it fed. If an input is estimated at prediction time, save that estimate. Then the error can be split afterwards into the part the hypothesis owns and the part it doesn’t.
Always compare against a simple persistence baseline.