Failure 06 — No optimization of the policy itself

We've optimized plans for thirty years. We've barely begun to optimize policies. A plan is a specific committed set of decisions — release this PO on this date in this quantity. A policy is the posture that produces plans — under this kind of demand shock, with this much lead-time variance, against this customer-priority hierarchy, here is the shape of the decision. Under uncertainty, the right question isn't which plan is optimal; it's which policy is.

Meta-optimization treats the policy as the thing being optimized. Two policies run against the same sampled uncertainty over the same horizon produce two distributions of outcomes. The dollar difference between those distributions is what better policy is worth — under the customer's own demand patterns, against the customer's own named risks, on the customer's own data. At MIC, the Pareto frontier across 280 nearby policy variants identifies a knee point three positions away from the current operating policy. Conservative-v3 — the policy currently active — gives up $11.8M of expected EVA relative to Balanced-v17 (the knee) for 1.8 service points. Lean gives up $1.4M relative to Balanced for $14.6M of working-capital release. The CSCO and CFO stop arguing values and start arguing numbers.

Why this hasn't been done at scale: meta-optimization is computationally demanding (200 candidate policies × 100 iterations each = 20,000 underlying solves to evaluate the frontier), and the ranking-and-selection math that makes it tractable in a weekly batch has matured only in the last decade. The compute is there now. The math is there now. What's been missing is a production system that does it end-to-end. Chapter 4.4 develops VYAN's approach.