The Case for Extraction Over Rewrite

The instinct when a codebase feels illegible is to rewrite it. Clear it out, start clean, do it properly this time. Chapter 5 argues that instinct is usually wrong, and not for sentimental reasons. A rewrite consumes months of your most senior engineers' time, and at the end of it you have cleaner code that still doesn't declare why it works the way it does. The tribal knowledge that made the old system risky just moves into the new one, unwritten, waiting for the next person who has to guess at it.

Extraction is a different bet. Instead of replacing a system, you pull one capability at a time into a contract that says what it does, what it needs, what it promises, and what it explicitly doesn't do. Each extraction ships independently, each one is immediately useful on its own, and the value compounds as the capability graph fills in. A rewrite pays off once, if it pays off at all. Extraction pays off capability by capability, starting with the first one you finish.

The harder question is which capability to extract first, and that's where the priority model earns its keep. Three axes decide it: business criticality, how much revenue or compliance risk actually sits inside the capability, tribal knowledge density, pulled straight from the legibility audit score, and agent touchpoint frequency, how often agents are actually going to work inside this capability once they're operating in the system.

Those three axes get multiplied, never added. That choice matters more than it looks. A capability that's mission-critical and completely opaque but that nobody, human or agent, ever touches shouldn't jump the queue ahead of one that's moderately risky and gets hit constantly. Multiplying keeps a zero on any single axis from letting a capability sneak to the front, because a zero anywhere means it isn't a real candidate yet, no matter how bad the other two numbers look.

The result sorts into three phases. A perfect score of nine lands in phase one: extract now. Scores from four to eight are phase two: next in line. Scores of one to three are phase three: worth tracking, not urgent. A score of zero means the capability isn't a current extraction candidate at all, and knowing that is useful too. Not everything needs a contract on day one, and pretending otherwise is how extraction programs stall under their own scope.