Failure 02 — Stale master data dressed up as single-value assumptions

MIC's planning system carries a 14-day lead time for the Taiwan FPGA component lane into Penang. The number was set in 2023 by an analyst who left in early 2024. Nobody has touched it since. Looking at the actual receipts for that lane over the last eighteen months: mean 14 days, standard deviation 4 days, 85th percentile 19 days, 95th percentile 24, 99th percentile 31. The single number — 14 — represents the statistical mean of a distribution that runs 9 to 31. The system commits as if the mean were the whole truth.

Downstream, planners learn not to trust the master data. They buffer — at the SKU level, at the location level, at the customer level — to absorb shocks the master data won't tell them are coming. The buffers trap working capital in inventory nobody can defend at the next CFO review. And the buffers don't actually fix the problem, because they're sized for average drift, not tail drift. The lane that ran 24 days last month broke through every buffer in the network. The problem isn't that the master data is wrong; the problem is that it's a single value in the first place.

VYAN's answer: give up on static parameters entirely. Every input driver in VYAN is a learned distribution fitted from the customer's own transaction history — lead times, capacities, scrap rates, customer cancellation rates, FX, commodity prices, supplier promise reliability. The Decision Policy chooses, explicitly, how much of each shape to plan against, in named percentiles. The whole shape is always available; the commit is a deliberate choice from it.

Chapter 2.2 develops the distributional shift; node 1.3.4 develops the executive choice that picks the percentile.