Complexity Must Pay Rent

Work from the bottom of the ladder. Climb only when a lower rung cannot settle a load-bearing claim.

A count, a slice, a query. Is the damage concentrated? Is the current rule even being followed? This is where the ranking often dies or lives. A Pareto chart is not unsophisticated. It is frequently the whole claim.

A back-of-envelope economic model. What does a miss cost? What does a unit of inventory cost? Are those two numbers in the same universe? If they are not, the answer may already be plausible before anyone fits a demand model.

A simple statistical relationship. Is there something stable enough to use, and does it change an alternative? A correlation is not a stocking rule. A stable split of items into classes might be.

An experiment. When you can actually intervene and what you would learn is decision-relevant. Significance is not the decision; rollout is. Climb here when a cheap observational cut cannot settle a claim that a two-week trial could.

A forecast. As an input to a rule, never as the product. A forecast that does not change an order is a diagnostic with a press release.

A simulation. When the system’s dynamics matter, when queues, bullwhips, and constraints bite in combination, and when averages lie. Not when you wanted something impressive for slide nineteen.

An optimization model. When the alternatives are allocations of scarce resources under real constraints, and a heuristic would leave meaningful value on the table in a way you can estimate.

The rent

Each rung costs something real: build time, review time, and, most expensively, legibility. A planner who does not understand the rule will route around it, which means a model nobody can follow has a lower effective value than a worse model that gets used.

So the test for climbing is specific. Name the claim the current rung cannot settle. If you cannot name one, you are climbing because the higher rung is more interesting to build, which is an honest motive and a bad reason.

Heuristics are first-class

This is the part practitioners resist hardest, and the resistance is about status rather than about performance.

A heuristic that a planner will actually follow, that degrades gracefully when an input is missing, and that can be explained in one sentence is frequently the better decision even where a solver would produce a tighter number. The comparison that matters is not model against model. It is the realized outcome of a rule people run against the realized outcome of a rule people work around.

Get the answer a second way

Whatever rung you land on, reach the answer once more by a different route, and cheaply.

If the optimization says class-based stocking saves 1.2 million dollars, check whether a back-of-envelope on the protected class lands in the same order of magnitude. When the two disagree, one of them is wrong and you have found it before the meeting rather than during it.

This is the single highest-return twenty minutes in most projects, and it is skipped almost universally because the first answer is already in hand and checking it feels like doubting yourself.