Deriving with the machine: principle-first learning as a firm capability and a safety property

Generative AI lowers the cost of producing analysis, and early evidence shows it can also lower the judgment of the people who use it. People check AI output less as the tool improves, and learners who receive direct answers perform worse once the tool is removed. This note proposes a practice that keeps judgment growing while the machine takes on more of the execution: derive the principle behind an output, encode it as a written rule, and reuse the rule as an instruction on later tasks. The note sets out the mechanism. A person who traces an output to its 1st invalid step stores the correction twice: once in memory, and once in a written rule that applies on every later task and every model. When a team runs the practice, its rules become calibration capital: judgment the firm owns, can audit and can carry from one model to the next. The note states 4 testable propositions and outlines 2 tests any firm or research group can run, 1 on people and 1 on rule libraries. The practice also bears on AI safety. The internal reasoning of current AI systems cannot be read in full, so a person's ability to reconstruct and challenge an output is the only check available at the point of use. How people are taught to work with AI decides how many such people exist.