Design the Rule

A challenger that wins the ranking still has to displace a rule that is already running, and that rule holds advantages no analysis removes on its own. The incumbent occupies the default, costs nothing to follow, has defenders whose scorecards depend on it, and needs nobody to argue for it. Designing the rule means arranging for those advantages to pass to the challenger.

The work has two halves. The first is anatomy: naming the five parts the rule needs, so that none is left for the floor to improvise. The second is behavior: arranging the default, the exception path, and the feedback so that following the new rule is the path of least effort.

The state. What is visible at the moment the rule fires. Not what was true in the historical extract, but what a person or a system can actually see on Tuesday morning.

The decision. What the rule is allowed to do now. A decision is an action available in the current state, not a hope about an outcome.

The information that will arrive. What you do not control and do not yet know. If the rule pretends this is empty, you have written a deterministic plan and called it a policy.

The transition. How the state, the decision, and the new information combine into the state the rule sees next time. It is the reason a rule that looks good in a snapshot oscillates within a week.

The objective. What you are accumulating across those transitions, in units of consequence rather than in house metrics. A rule with no objective gets run against whoever’s scoreboard is loudest that month.

Most analysis specifies two of the five

The decision is whatever the model outputs, and the objective is whatever it was optimized on. The state is whatever was in the extract, the incoming information is a point forecast treated as fact, and the transition is a backtest loop that does not contain the people who will route around the rule. Rank alternatives on its two parts and you will pick the one that fits history; install it and you will discover the other three, at the operator’s expense.

The state is not the database

The state is what the rule can see in forty seconds, where it actually lives. Walk the rule through its inputs and ask of each: is it known at the moment the rule fires, and is it known where the rule fires? A quantity that lives only in finance is absent everywhere the rule runs, and a rule that consumes it will be run on a guess.

The most common failure is hindsight. At Meridian Motors the natural classification is by whether a part strands a vehicle, which is perfectly observable in last year’s data and unknown when the rule fires. The repair has two steps. Find something visible now that stands in for it, such as appearance in repair orders that held a bay. Then re-rank the alternatives using the proxy, which will misclassify some parts. Skip the second step and a clean version of the challenger has been ranked against the incumbent’s actual performance, and it has won for reasons that do not survive April.

Write into the rule what it does when it is unsure, and rank the alternatives on how they perform degraded. If planners consult the unofficial spreadsheet at the moment of action, it is part of the state. Either the challenger absorbs what it does, or it remains the rule and the challenger becomes a second rule layered on top.

Information you do not control is not a residual

A rule that is optimal on a point forecast is a rule for a world that will not occur. And some rules are valuable because they change which information arrives: letting a supplier see actual consumption rather than orders is not a forecasting improvement but a different rule, and it belongs on the menu as one.

The transition is why the problem comes back

You order less on Monday, and the shelf does not know until the pipeline empties. During the lag the shortage is still visible, so the rule orders less again, or somebody overrides it. You do not need a simulation of the firm, only one honest sentence about how today’s action shows up in next week’s state, including the delay. “Classes recompute monthly from twelve-month traces, so a new platform will be misclassified for up to a year” is a transition, and it belongs in the ranking rather than in a footnote.

Make the right action the default

Whatever the screen prefills is the policy for every case nobody examines closely, which is most of them. A challenger that leaves the incumbent’s default in place asks every operator to override the screen on every line, and it will lose that contest whatever its merits.

At Meridian Motors the class-based quantity is the number in the box. The expedite screen opens with the class’s path already selected.

Where the rule is least reliable, such as new items, the right default may be an empty box that requires entry.

Make an exception cost a sentence

A rule that allows no exceptions will be bypassed, out of the owner’s sight. A rule whose exceptions cost nothing will become its exceptions. One written sentence stating what is different about this case is usually the right price: cheap enough that genuine exceptions still happen, expensive enough that habit stops routing through it, and legible enough that a week of sentences can be read in a few minutes.

At Meridian Motors a protected-class expedite requires a one-line reason and is charged to the requesting region, and every exception appears on a weekly list Priya Shah reads. A part excepted three times in a quarter goes to reclassification review, so repeated exceptions change the rule instead of quietly replacing it.

Close the loop

A rule improves only if the people running it can see what it does, soon enough to change the next action. Today the planner who triggers an expedite never sees its cost. Closing the loop means routing each expedite’s cost back to the requester on the weekly list, and putting the rule’s scoreboard in front of the planners who set the quantities.

Write the five, then the three, before you specify screens

For the class-based rule: the state is on-hand, on-order, current class, and an open-expedite flag. The decision is order quantity by class and whether to approve an expedite at a named price. The information is demand, dealer bypass, supplier misses, and new platform launches. The transition moves on-hand with receipts and demand and recomputes classes monthly from a visible proxy. The objective is holding dollars and stranded-job cost by class over two quarters. Then the default, the exception price, and the feedback.

The status quo deserves the same sentences, and filling them is often the moment the current rule turns out to be an unpriced exception path.

A five-minute test

Write the five parts and the three settings for your leading alternative, one sentence each. For every input in the state, mark whether it exists when and where the rule fires. If a row is blank, or the default on the screen is still the incumbent’s output, you have a slogan about a rule rather than a rule.

Exhibit 1A repeating rule is a loop, and the floor fills whatever part you leave unnamed