I built Code-role after two real projects showed the same problem at different stages. In DeepBrain, 1,750 passing tests, S50 at 50/50, LongMemEval-S at 499/500, and 100/100 source joins still did not prove the product milestone. Independent Evaluation kept the gate at 0 because fair comparison, raw reruns, clean reproduction, and cost/SLO proof were missing. In Leaper Agent, the Project Manager rejected a professional-looking evaluation baseline before Engineering started: task artifacts were missing, holdout was exposed, and grader/runtime conditions were placeholders. The correction had to produce real tasks, physical holdout separation, calibration, command contracts, and integrity hashes. Code-role is the control loop that came out of those failures. It is local-first and open source. The default profile has four workstations: Project Manager, Product Strategy, Engineering, and Independent Evaluation. I would value feedback on where this control is useful and where it still adds too much process.

Code-role

Keep AI coding aligned to one software milestone

Keep AI coding aligned to one software milestone