The cost of independent validation

Last week I called the margin period of risk an agreed number, not a computed one. Here is the harder question underneath that: agreed once, and left alone for how long?

Ten, five, twenty days. Nobody revisits those floors often, because revisiting them means independent validation, not recalculation. Independent validation is slow and manual, a function that did not build the model checking it well enough to argue with it, then repeating that on a schedule for as long as the model stays live. Adjoint algorithmic differentiation never touched that cost. It was never a computation problem.

What changes if checking itself gets engineered rather than performed by hand: an assumption declared with a test attached, checked again by an independent process computing the same number a different way before anything ships. None of that is a model predicting markets. It is the discipline that makes checking cheap enough to repeat often, not just once at launch.

A floor fixed for years because revalidating it was expensive starts to look like a choice made under a cost constraint that does not have to hold once checking is engineered this way instead of done by hand. The margin period of risk is not obsolete. The cost structure that froze it in place might be.

This first appeared on LinkedIn on 5 August 2026. If you want to comment or discuss, that's the place.