Cut the Question Into Parts

A tree cuts a question into parts you can work on separately. Two properties make a cut usable, and each fails in its own way.

No overlap. When branches overlap, evidence gets double-counted and you cannot tell how much weight any one factor carries.

No gap. When something relevant falls outside the set, the analysis is not merely incomplete; it is systematically biased toward whatever happens to be inside the frame.

Where each property actually fails

Overlap failures are irritating and rarely fatal. You notice them when the same number appears under two headings, and you fix them by restating the cut.

Gap failures are expensive and nearly invisible, because nothing in the tree points at what is missing. The recurring gap in technical work is the null branch: a tree that decomposes how to change a rule, with no branch for leaving it alone. A document justifying its own existence rarely concludes that its subject did not need addressing, and that structural bias is worth naming out loud.

The cheap defense is a single sentence for every published cut, naming the branch you deliberately excluded and why. It costs almost nothing and it is the highest-yield habit here.

Choosing the cut

Most questions admit several cuts and they are not equally useful.

Cuts that are true by construction carry their own proof. An arithmetic identity holds: holding cost is units multiplied by unit cost multiplied by time. A binary split holds: the rule was followed or it was not. You can nest these indefinitely without losing the property, though you will lose readability first.

Cuts that split on a single dimension are workable. The common structural error is a first level that mixes dimensions, such as branches for “forecasting,” “the replenishment rule,” “dealer behavior,” and “seasonality.” That mixes a system-component split with a causal-class split, and it will leak immediately. Split on components, or split on causal classes, and put the other dimension at the second level.

Prefer cuts where each branch can be bounded with a number, even roughly. A branch you can bound can be eliminated cheaply. A branch defined qualitatively will absorb effort indefinitely, because nothing ever conclusively disproves it.

Prune, then go

The value of the tree is not the tree. It is the order in which you work the branches, and the branches you never work at all.

Before investigating anything, mark every branch that cannot change the ranking of your alternatives. Those branches are done. Write the sentence explaining why you are leaving them alone, because you will need it when someone asks in week five, and because writing it is how you check that the reasoning holds.

For what remains, put two rough numbers on each branch: how likely it is to matter, and what it costs to check. Work in descending order of likelihood divided by cost. This will often tell you to check a cheap, improbable branch before an expensive, likely one, which feels wrong and is right.

What reaches the write-up

Almost none of it. The reader does not want your tree. They want the recommendation, the reasons, and one line confirming that the obvious alternatives were eliminated. An appendix can hold the rest for the one reader in twenty who wants it.