Delta PVA — the policy-comparison primitive

Delta PVA is the dollarized difference between any two policies, evaluated on the same data with the same horizon and the same joint sample. The primitive that makes policy comparison rigorous instead of rhetorical.

Three principal use cases anchor it. Migration decision — should we move from the current operating policy to the recommended knee policy? Delta PVA between them is the answer. A/B comparison — should we run Conservative or Lean for the next quarter? Delta PVA across the two postures, by KPI family, settles it with attribution. Engine arbitration — when VYAN orchestrates multiple solvers (chapter 3.6's engine tournament), Delta PVA between the solvers' output policies names the winner.

The confidence interval is what makes Delta PVA load-bearing for board narratives. Delta PVA is computed across N joint-sample iterations, which gives it a distribution. The reported number is the expected delta; the confidence interval comes from the iteration variance. A Delta PVA of $11.8M with a 95% CI of [$8.4M, $15.2M] is defensible — the lower bound is still a substantial number, the upper bound is realistic, the system is telling you with confidence what's true. A Delta PVA of $11.8M with a 95% CI of [-$2M, $25M] is not yet defensible, and the system tells you that too — N is too small, the variance is too high, you need more iterations before you can sign the migration paperwork.

Delta PVA is the measurement framework that makes the engine tournament (3.6) work, and it's the metric the workshop bootcamp (chapter 8.3) produces on the customer's own data. The same primitive that arbitrates between SAP IBP and Kinaxis at the start of an engagement also arbitrates between Balanced-v17 and Lean three quarters later.