There are no side effects

You shortened payment terms to protect cash. Cash improved. Two quarters later your best distributor is ordering less, because your terms are now worse than the competitor’s, and nobody connected the two.

That was not a side effect. Side effects are not a feature of reality. They are a report on the boundary you drew.

Why these three models

The decision is how wide to draw the boundary before you act, and how to run the post-mortem when something unexpected arrives. The features that fire are a trajectory nobody has plotted, a measure that has started being optimised, and an estimate built from your own case rather than from a class.

Three lenses. Behaviour modes produce a cycle answer about what shape you are actually looking at. Goodhart produces a complex answer about the specific way measurement redraws the boundary from underneath you. The inside view produces an equilibrium answer about why your boundary defaults too narrow in the first place. The second and third are the two commonest causes; the first is how you tell them apart.

1. The reframe, and it is not a word game

Sterman puts it plainly and it is worth quoting because the phrasing does the work: we frequently talk about side effects as if they were a feature of reality, and they are not. In reality there are no side effects, there are just effects. The ones we thought of in advance, or that were beneficial, we call the main or intended effects. The effects we did not anticipate, the ones that fed back to undercut our policy, those are the ones we claim to be side effects. Side effects are not a feature of reality but a sign that our understanding of the system is narrow and flawed.¹

The reason this matters operationally rather than philosophically is what it does to a post-mortem. “There was an unforeseen side effect” closes the inquiry. “My boundary excluded the distributor’s alternatives” opens it, and names something you can fix.

The underlying cause is that we act as though cause and effect are close together in time and space, when in complex systems they are often distant in both.¹ The distributor’s response was two quarters and one relationship away from your terms decision, which is exactly far enough to look unrelated.

So the discipline is to draw the boundary explicitly, before acting, and to draw it wider than is comfortable. That is not a call to consider everything, which is impossible. It is a claim about where to spend a fixed budget of attention, and the guidance from the same source is unambiguous: a broad model boundary is more important than a great deal of detail.¹

2. Name the shape before you explain it

Before diagnosing any unexpected consequence, plot it. This step is skipped almost universally and it is the cheapest one available.

There is a small set of fundamental shapes and each implies a different structure underneath.¹ Exponential growth means a reinforcing loop with nothing balancing it. Goal seeking means a balancing loop already arrived. Oscillation means a delay between an action and the information about it. S-shaped growth that turns over into decline means overshoot, and overshoot is the one to catch early because the correction is not gentle.

Naming the shape narrows the search before you pick a cause. Diagnosing a boundary error on a trajectory that is actually a one-off event produces an elegant analysis of nothing, and the reverse produces an event explanation for a structural problem that will simply recur.

Behaviour modes, the shape lens

• Assumes: a trajectory belongs to a small set of shapes, each implying a different feedback structure.
• Fits because: something changed and you are about to explain it from memory.
• Breaks when: the series is short or noisy, which at small volumes it usually is.
• Evidence: grade B. Descriptive and useful, with no strong predictive claim attached.
• Counteracts: explaining a pattern before establishing that it is one.
• May reinforce: reading structure into three noisy points.

3. The boundary error that measurement creates

The second lens is the most common single cause, and it is the one that redraws your boundary without your involvement.

A measure that becomes a target stops measuring what it measured. The mechanism is not mysterious: the people inside your boundary optimise the number, and the part of the system that was silently holding the number honest sits outside the boundary, unmodelled.

Meadows names the same failure from the other direction and supplies the exits. Systems have a terrible tendency to produce exactly and only what you ask them to produce, so if the goal is defined badly, if it does not measure what it is supposed to measure, the system cannot produce a desirable result.² And where rules generate behaviour that gives the appearance of obeying them while distorting the system, that is rule beating, whose exit is to redesign the rules so creativity runs toward their purpose rather than around their wording.²

The practical test before you introduce any metric is one question. If someone optimised this number hard, without malice, what would they stop doing? Whatever that is, was outside your boundary, and it is where your side effect will come from.

Goodhart, the measurement lens

• Assumes: the measure becomes a target and behaviour reorganises around it.
• Fits because: the unexpected consequence followed a change in what you measured or rewarded.
• Breaks when: nobody inside the boundary can actually influence the number, in which case the mechanism has nothing to work with.
• Evidence: grade B plus. Widely observed with named exits, and not formally tested as a set.
• Counteracts: adding a metric and expecting the rest of the system to hold still.
• May reinforce: measuring nothing, which is worse.

4. Why your boundary defaults narrow

The third lens explains the bias itself, and it is not carelessness.

Left alone, we build an estimate from the features of the case in front of us: our team, our customers, our decision. That inside view feels like evidence because it is specific and detailed, and its boundary is drawn exactly at the edge of what we can see clearly. Everything beyond that edge is not consciously excluded. It is simply never considered, which is why the exclusion leaves no trace to notice later.³

That is why the fix has to be procedural rather than attitudinal. Asking people to think more broadly does not work, because the boundary was never a decision they made. Asking them to write down what is outside the line does work, because it converts an absence into a list.

Write down what you left out. That list is what you have agreed to be surprised by.

Inside view, the boundary-default lens

• Assumes: estimates built from case features are narrower and more confident than the evidence supports.
• Fits because: the excluded actor was not rejected, they were never considered.
• Breaks when: the case genuinely is unusual and the wider class misleads, which happens more often than base-rate advocates concede.
• Evidence: grade B. Well established in direction, with context-dependent magnitude.
• Counteracts: a boundary drawn at the edge of what is visible.
• May reinforce: boundary sprawl, where everything is in scope and nothing is decided.

The levers, cheapest first

• Ban the phrase. Not for tidiness. “Side effect” ends an inquiry and “boundary error” continues it, and the difference in what gets found is large.
• Write the exclusion list before acting. Three lines. Who is outside this analysis and might react. It takes minutes and it is the whole intervention.
• Plot before you explain. Ten minutes with the actual series, before anyone offers a cause.
• Run the optimisation question on every new metric. If someone hit this number hard without malice, what would they stop doing?
• Name the actor, not the effect. Every unexpected consequence has somebody who responded. Find the person, not the phenomenon.
• Widen before you deepen. When you have a fixed hour, spend it on who else is in the system rather than on more precision about who you already listed.

What to do this week

Do now, sized at fifteen minutes, effect immediate. Take the last decision that produced an unexpected consequence and write down who was outside the boundary you drew at the time. Reversible, free, and dominant across every scenario about whether it was foreseeable.

Hedge, where the premium is the whole loss, live before the next decision. Add one line to whatever template your team uses for proposals: who is outside this analysis and might react. If your boundaries were always wide enough you have added a line nobody needed, and that is the entire downside.

Defer and trigger, size fixed now. Do not re-examine every past decision. Pre-commit the trigger: the next time someone says the words side effect in a review, that meeting stops and the exclusion list gets written before the discussion continues. Decide now who is allowed to call that halt, because a rule with no enforcer is a preference.

Watch the arrivals as well as the dates. The exclusion list lands immediately. Whether it improves decisions cannot be seen until the next unexpected consequence fails to arrive, which is a non-event and therefore never celebrated.

What usually happens next

Run the break test first. Has a rule changed, has an actor entered or left, has a measurement become a target? The third is the specific one to check here, because a new metric changes the process you are extrapolating from, and the historical series is then describing a different system.

If nothing broke, the pattern is reliable and slightly bleak. The unexpected consequence gets attributed to an external factor, the original policy is retained, a compensating policy is added on top, and the compensating policy generates its own excluded actor. Three layers later nobody can say why the process has the shape it does.

Subtract the counterfactual before crediting the fix. A consequence that stopped after you intervened may have stopped because the actor causing it changed strategy for their own reasons. The test is whether you can name the mechanism by which your fix reached them.

What this ensemble cannot see

All three lenses assume the excluded actor is findable once you look. Frequently they are not.

The boundary that matters is sometimes several steps out: your customer’s customer, a regulator who has not acted yet, a competitor’s unannounced decision. Drawing the boundary wider is genuinely good advice and it has no natural stopping point, and this framework offers no principled place to stop. In practice people stop where their patience runs out, which is not a criterion.

There is a sharper version of the problem. Widening the boundary has a real cost in speed, and a company that models the whole ecosystem before every decision will lose to one that acts and absorbs the consequences. Nothing here prices that trade, and the article’s advice pushes in one direction only.

And one property none of these models contains: banning the phrase changes what people say before it changes what they think. You will get boundary language covering the same shrug, and it takes a while to tell the difference.

The one action that survives the ignorance: at the next review of something that went unexpectedly, ask one question before any explanation is offered. Who responded to what we did? If nobody in the room can name a person, the analysis has not started, and the answer that follows will be a story about the world rather than a finding about your boundary.

Who has to move

The person who needs this is whoever runs the post-mortem, and the phrase is comfortable precisely because it distributes blame to nobody. The cheapest first test is the three-line exclusion list on the next proposal. If the actor who eventually reacts is on that list, the boundary worked. If they are not, you have learned exactly where your default edge sits, which is more useful than any single fix.

Sources and notes

1. John D. Sterman, Business Dynamics: Systems Thinking and Modeling for a Complex World, McGraw-Hill. The passage on side effects, that we talk about them as if they were a feature of reality when in reality there are no side effects, only effects, and that they signal a narrow and flawed understanding of the system, is chapter 1, section 1.1.2 on policy resistance. The related point that we act as if cause and effect are closely linked in time and space when in complex systems they are often distant in both is in the same section. Problem articulation and boundary selection as the first step of the modelling process, and the principle that a broad model boundary is more important than a great deal of detail, are chapters 3 and 2 respectively. The fundamental modes of dynamic behaviour used in section 2 are chapter 4.
2. Donella H. Meadows, Thinking in Systems: A Primer, Chelsea Green. Seeking the wrong goal, including the observation that systems have a terrible tendency to produce exactly and only what you ask them to produce and that a badly defined goal cannot produce a desirable result, is chapter 5. Rule beating, defined as perverse behaviour that gives the appearance of obeying the rules while distorting the system, with the stated exit of redesigning rules to release creativity toward their purpose rather than around them, is in the same chapter. Note that Meadows presents the system traps as widely observed archetypes rather than as a formally tested set, and they are used here accordingly.
3. Matthew A. Cronin, Cleotilde Gonzalez and John D. Sterman, Why don’t well-educated adults understand accumulation? A challenge to researchers, educators, and citizens, Organizational Behavior and Human Decision Processes 108(1), 2009. Author copy: https://www.mit.edu/~jsterman/CroninGonzalezSterman061210.pdf. Used for the claim underlying section 4, that the failure to reason correctly about system structure is not attributable to graph literacy, contextual knowledge, motivation or cognitive capacity, which is why the boundary fix has to be procedural rather than a matter of thinking harder.

A note on the strongest objection to this piece. Widening the boundary has no natural stopping point, and an organisation that takes this advice literally will model itself into paralysis. The honest position is that the default error runs narrow, so the correction runs wide, and neither this article nor its sources can tell you where to stop. The exclusion list is useful because it is finite and written, not because it is complete.

Joshua Agonya Pi’Rwot, Founder.