The book began with a bug that could be fixed twice. One repair changed the program. The other changed what the development process would notice the next time. That modest difference opened onto a larger problem. Once the machinery of improvement can itself be changed, the rules, evidence, and people judging the change are no longer safely outside the loop.
A horizon is a direction, not a score. Capability helps us travel; judgment still chooses what is worth moving toward. Photo by Alex Gruber on Unsplash
Software made the loop easy to see. A failure can become a test that runs tomorrow. Version control preserves the old state, and rollback keeps one bad change from becoming permanent. The same pattern becomes less legible in a person, where the observer has pride and blind spots, and still less legible in an organization, where incentives shape which failures are reported at all.
Markets, politics, evolution, and competition removed the comfort of a fixed environment. Other people learn too. They respond to a price, a rule, a defense, or an adaptation, and their response changes the next round. Improvement becomes relational. A participant can become more capable while the shared game becomes more wasteful or cruel.
Those comparisons leave three questions worth keeping. What changed? What evidence says it was better? What will carry forward into the next attempt? The questions are simple enough to use on a study routine, a company process, or an AI system. Honest answers are often difficult because the people answering them helped choose the objective and may benefit from the result.
Recursive improvement increases the need for judgment because the system is changing how later answers will be found and believed.
The safeguards developed in the previous chapter are forms of practical humility. Use more than one kind of evidence. Keep authority narrower than capability. Make changes visible. Preserve a way back. Let someone outside the immediate loop object, and make sure the objection can still stop the next turn.
None of this supplies a final definition of the good. Efficiency, prediction, scale, and control are powers. They can support health, learning, friendship, craftsmanship, public order, and other goods, but they cannot choose among them. A system may become extraordinarily good at reaching the target placed before it. The tighter group does not prove the target was worth choosing.
The human boundary: Friendship, courage, faith, care, and craftsmanship grow through practice and correction. Treating them only as scores to maximize can damage the goods the measurements were meant to serve.
The best improvement processes remain open to truths they did not generate. They preserve experience without making old rules unquestionable. They allow experiments without asking other people to bear unlimited risk. They enlarge human agency while leaving human beings responsible for where that agency is directed.
A better test suite, study method, company, market, constitution, or learning machine can give people more room to act. That is a real good. It is still a means. The final judgment belongs with the people who must decide what kind of life the improved machinery is helping them build.