Part III: The Craft
What the Judgment Is For
Part II built a practitioner, and it left a paper trail. That was not incidental.
Look at what you hold if you did the work. The loop templates, with your role splits written down so the producer never approves its own output. The model dossier, one profile per actor you cast, with the battery that produced it and a re-run date for the next upgrade. The configuration file, your standards externalized in your own hand, the one part of your judgment that survives a change of tools. And the first pages of the proficiency record: the commits written before the model answered, the first catch rate, the ledger with its open entries waiting for ground truth. Four artifacts, each one a side effect of practice, each one dated, each one yours and defensible line by line.
Hold the stack for a moment, because the rest of this book keeps adding to it, and by the final chapter the stack will turn out to have been the point.
A maintained capability that never gets deployed is a fitness routine for a race you do not run. So this part deploys it, at the three sites where the product manager’s judgment cannot be handed to anyone else, not to the model, not to the eval owner, not to the engineering lead who owns the runtime.
The first site is the brief: before the agent exists, someone decides what it is for, where its authority ends, and who answers when it is wrong. That chapter adds the two-brief rubric to your stack. The second site is the gate: someone reads the eval and decides whether the green checkmark is a verdict or a document with gaps. That chapter adds the gate owner’s checklist, earned by building one small suite with your own hands. Between the gate and the room sit two chapters that serve all three sites: the steady-state read of your running agent, which adds a weekly page that turns production into roadmap, and the audit of your own toolbox, which adds nothing new to the stack but sharpens everything already in it. The third site is the room: carrying all of this into an organization calibrated to ship without becoming the person the room routes around. That chapter adds the post-mortem template and the dissent ledger, the record turned briefly on your own objections.
Part II asked whether you can still see. Part III is where seeing has consequences, and where the record of it grows.