Most decisions are several decisions

The household sends one sentence: plan dinners for four this week, Tuesday and Thursday are busy, use the spinach before Wednesday, avoid meals we ate last week. The ticket that comes back reads "meal planning," estimated at one sprint, owned by one prompt.

A team that has learned the open-closed sort will try to apply it to that ticket and find there is nothing to apply it to. The sort asks whether competent people would agree on the answer and whether the rule behind it could be written down, and both questions need a single decision to be asked about. "Meal planning" is not a decision. Behind it sit interpretive work, exact arithmetic, contextual comparison, a question only the household can answer, and an authorised purchase, which means the sort returns a different verdict at every point. The skill that decides the architecture is therefore not the sorting but the cut that precedes it, because until one request has been separated into the decisions it was concealing there is nothing to classify, nothing to evaluate, and nobody to hold responsible.

Helpful context: Open, closed, and closing supplies the two sorting questions this page assumes. Where does a decision live? and judgment chains pick up what happens to the pieces once they exist. Mistakes both ways (next in this track) names the symmetric errors the cut prevents.

The sort has nothing to sort

Fused requests are the normal case rather than an occasional mess, and the reason matters because it tells you where to look for them. Work arrives fused because the closed part travelled alongside something that genuinely needed judgment, and whoever wrote the first template had no cause to pull them apart. The household said one thing, so the product answered with one thing. Nothing in the request advertises its internal structure, which is why the fused version survives review: it reads as a feature.

Fusion does not announce itself as a failure, and what it does instead is make later questions about the system unanswerable. Ask how accurate meal planning is, and the honest answer is a number mixing interpretation, arithmetic, and taste, so it cannot tell anyone what to improve. Ask who owns the choice to put a forty-minute braise on a busy Thursday, and the answer is a paragraph of a template, which cannot be held to anything. Both questions fail for the same reason, and it is not missing instrumentation: the thing being asked about was never identified, so there was nothing for a measurement or an owner to attach to.

Cutting one sentence into seven decisions

Performing the cut means walking the sentence from request to outcome, stopping at every point where an answer is produced, and asking the two sorting questions there rather than about the whole. Done on the household's sentence, using the Weekly Meal Companion terrain that runs through this collection, it produces seven answers, and they do not resemble each other.

Sub-decision	Character	Correct home
Interpret the current week	Contextual and evaluable	InterpretWeek within agent strategy
Identify confirmed allergies	Exact once knowledge is endorsed	Deterministic validation
Decide whether a candidate fits a busy evening	Fuzzy but strongly patterned	AssessMealPracticality
Compose the weekly plan	Multiple valid outcomes, competing goals	Agent-led composition
Clarify an ambiguous preference	Missing information owned by household	AskTargetedQuestion toward human
Calculate portions	Exact and verifiable	Deterministic code
Purchase groceries	Exact execution, financial consequence	Human approval, then deterministic action

Seven links and three kinds of home, out of one week of ordinary planning. The count is what the ticket concealed, and the spread of homes is what makes the count matter, because two of the seven belong in code, three want a cognitive component with an evaluation suite, one belongs to the household, and one waits for approval before money moves. A single template asked to produce all seven has been handed work from three different materials without anyone deciding that it should be.

The same shape appears wherever one control stands in for a decision. The invoice question from The Cognitive Unit, whether a variance can be approved automatically, hides date arithmetic that is closed, a credibility assessment that is open with a fact of the matter, and a question about how much tolerance the organisation extends to a vendor who keeps producing variances, which has no fact of the matter at all. Three materials behind one button, and the button gives no indication that the third question exists.

The cut starts with the closed parts

There is an order to the cut, and it is not the order the request suggests. You do not begin by asking what the judgment is, because the judgment is the part you cannot yet describe. You begin with the parts you can describe exactly, which are the closed ones, and lift them out. What remains once they are gone is the open decision, visible in outline for the first time, because the closed material packed around it was what made its boundary impossible to see.

That ordering also explains why fused components fail in a way that resists diagnosis. Ask one component both to judge whether Thursday can absorb a braise and to confirm that nobody in the household reacts to cashews, and it will occasionally get the cashews wrong, but the error arrives wearing the costume of a judgment error. Somebody then spends a week rewording the part of the template that talks about evenings, because that is where judgment problems are supposed to live.

Two neighbours own what happens after the cut. Closed decisions belong in code argues why the cashew check should never have been inside a judgment, and deterministic guards covers what to do with checks like it once they are out, since standing them in front of cognition rather than beside it changes how much reaches a model at all. The shape that recurs when the cut is made consistently, exact checks in, one open judgment in the middle, exact handling out, is argued at one cognitive unit, one open decision.

Find the closed parts and take them out of the cognitive unit. Not because purity is a virtue and not for tidiness.

The cognitive unit · Ch. 5

Fused and separated, side by side

Set the two architectures beside each other on the same week and the difference shows up in repair rather than in output. Fused, the product is one call that reads the request, invents candidates, argues the allergy in prose, decides Thursday can absorb a braise, and lays out the shopping list. None of those five behaviours has a name, so none of them has an address. Told that the Thursday choice was bad, the team can only edit the text that produces all five, and whether the edit disturbed the other four is discovered in production.

Separated, the same week runs through parts that can be pointed at. InterpretWeek returns a week representation the rest of the system reads, so the meaning of a busy Thursday is established once instead of re-inferred at every step. Deterministic validation settles the allergy before any candidate reaches the household, which means the exact part stops depending on how a sentence was phrased. AssessMealPracticality judges one evening at a time while ComposeWeek resolves the competing goals across all seven, so the practicality question and the arrangement question have different owners. Code handles portions and the list, and the purchase waits for approval. Told that the Thursday choice was bad, the team opens the one component whose entire responsibility is evening fit.

None of that required a better model, and what it did require is worth stating exactly, which is only that the parts have names. A regression in evening fit can be repaired without reopening interpretation or allergy enforcement, because the repair now has somewhere to land. The suite attached to evening fit produces a figure that moves only when evening fit moves, because nothing else is inside it. A run's bill can be attributed link by link, which is the arithmetic cost per decision depends on.

The decision nobody had authority to make

Localised failure and honest suites are the improvements teams expect from separation. There is one more consequence that nobody expects, and it changes what the product is permitted to do rather than how well it performs. Among the pieces the cut produces, there is usually at least one that no part of the system has the standing to answer.

In the invoice case it is the tolerance question: whether a vendor with four unexplained variances this quarter should be escalated or extended the benefit of the doubt. No amount of cognition supplies an answer to that, because it is not a question about the world. It asks what the organisation has decided to do, and when the organisation's own competent people disagree, the honest description is that nothing has been decided. Fused, the question was still being answered several times a day, differently on different days, by a template with no authority to answer it.

Separation does not resolve that question, and it is not supposed to. What it does instead is convert an invisible answer into a named outcome somebody has to deal with, which is uncomfortable and correct, because a system that surfaces an undecided policy is doing its job and a system that quietly invents one is accumulating a conversation nobody will enjoy having. The household version is milder and structurally identical, since interpreting an unfamiliar medical restriction is not a judgment the planner may make on the household's behalf however capable it becomes. Naming the link is what makes that refusal available as a designed behaviour rather than an omission.

What to do next

Separation is an analytic move, so the way to find out whether a product has had it performed is to try naming the pieces and see whether the names already exist. Take the request the system handles most often and count the answers it produces on the way to an outcome. Where the count exceeds the number of things in the architecture that have names, the difference is the part of the design currently being decided by prose.

Two things remain after the cut, and neither arrives with it. Each piece needs a single home, or one judgment ends up encoded in a template, a toggle, and a line of interface copy at the same time, which is the failure where does a decision live? treats. And the pieces are not independent of each other, because what one produces the next one assumes, which is the subject of judgment chains. For the two ways the cut goes wrong in the first place, closed work left inside cognition and open work forced into a growing tree of branches, see mistakes both ways.

Read next: Mistakes both ways.