Occam's Razor is usually introduced as a preference for simpler explanations.

Given two explanations that account for the same observations, prefer the simpler one.

That's useful as a first approximation, although it can easily turn into a preference for whichever story contains fewer words, fewer parts or less apparent complexity.

Reality has no obligation to be simple.

Sometimes the explanation with seventeen interacting variables is better because there really are seventeen relevant variables. Removing sixteen of them doesn't make the theory elegant. It makes it wrong.

So I've come to think about the razor somewhat differently.

Every claim should earn its keep.

Suppose I observe something and construct an explanation containing three assumptions. Then I add a fourth assumption that doesn't improve what the explanation predicts, distinguish between competing possibilities or account for evidence the first three couldn't handle.

What did the fourth assumption buy me?

Maybe nothing.

Now I have one more thing to explain, defend and carry forward without having increased the usefulness of the model.

That's the kind of complexity I want to cut.

This doesn't mean assumptions are liabilities merely because they're assumptions. Every useful model leaves things out and introduces some structure for relating what remains. The question is whether that structure helps us do something we couldn't do without it.

Sometimes additional complexity earns its place immediately.

A mechanic hears an engine making an unfamiliar noise. "Something is broken" is a wonderfully simple explanation and almost useless for repairing the car. Distinguishing among ignition, fuel, lubrication, timing and mechanical wear makes the explanation considerably more complicated.

The complexity buys diagnostic power.

That's different from adding an elaborate story about why the car chose this particular morning to betray us.

Both stories contain additional structure.

Only one helps us distinguish among possibilities we might actually encounter.

This is where Occam's Razor becomes more interesting to me than a generic celebration of simplicity.

The razor doesn't tell us how complicated the world is.

It disciplines what we're willing to claim about it.

If two models explain the available evidence equally well and one requires several additional commitments that currently do no explanatory work, I have less to lose by carrying the smaller model for now.

The words for now matter.

Tomorrow may provide evidence that requires another distinction.

Then the more complicated model may become the simpler adequate explanation, not because it contains fewer parts, but because removing those parts would now require us to ignore something we've encountered.

That makes simplicity dependent on the question we're asking.

Consider a map.

A road atlas leaves out property boundaries, elevation contours, sewer lines and the location of every tree. That's not because those things aren't real. They aren't necessary for the work that particular map is trying to do.

Give the same map to someone repairing a water main and its simplicity becomes a defect.

They need another representation.

Neither map is the true map.

Each preserves distinctions relevant to a particular purpose.

I think something similar happens when explanations become identities.

A person notices that they repeatedly respond to certain situations in similar ways and eventually finds a description that helps make sense of the pattern.

That description can be useful.

Then more claims accumulate around it.

Because I'm this kind of person, I must prefer these things. People like me behave this way. This experience means that. Someone who challenges the description must not understand me.

The original compression may have clarified something real.

The additions may or may not.

Occam's Razor doesn't tell us to abandon identity or reduce a person to the smallest possible description. A human being isn't improved by having fewer characteristics.

It gives us a question to ask about the explanation:

What does this additional claim allow us to understand?

If the answer is nothing—if the claim exists mainly to protect the previous claims—then perhaps we've added structure that no longer earns its keep.

Institutions can accumulate complexity the same way.

A procedure solves a problem.

An exception handles an unusual case.

Another rule prevents abuse of the exception.

A reporting requirement verifies compliance with the new rule.

A committee is created to review the reports.

Twenty years later, everyone is doing quite a lot of work, and nobody is entirely sure which original failure all of it is preventing.

That doesn't mean we should delete the committee.

Maybe the old problem remains real and the accumulated machinery is exactly why nobody has encountered it recently.

We have to investigate.

That's the important part.

Complexity isn't guilty until proven innocent.

Neither does longevity prove necessity.

We recover the problem the structure was meant to solve and ask whether the structure still solves it, whether the problem still exists and whether a less costly arrangement could now do the same work.

Sometimes the old rule survives the examination.

Now we've learned something.

Sometimes half the procedure disappears.

We've learned something then too.

Software makes this unusually visible.

A mature codebase accumulates dependencies, abstractions, compatibility layers, configuration options and special cases. Some of them represent unnecessary complexity. Others are compressed histories of failures nobody wants to rediscover.

Deleting code because the new implementation looks cleaner can therefore be just as foolish as preserving every line because somebody once thought it necessary.

The useful question isn't how much complexity exists.

It's what each piece is buying.

That changes how I think about my own explanations too.

I have a tendency to notice patterns across different domains. Something I learn while working on software appears to resemble something happening in an institution, a relationship or an economic system.

Sometimes carrying the pattern across domains produces a useful question.

That's enough.

I don't also need to claim that both domains are manifestations of the same underlying law.

The additional metaphysical claim may be true.

It may also contribute nothing to the observation that made the analogy useful.

If I don't need it yet, I can leave it out.

That isn't an assertion that reality contains no deeper unity.

It's a decision not to spend explanatory commitments before I have evidence that requires them.

This is where restraint becomes useful.

Not all restraint.

A theory so cautious that it refuses to make predictions can't be wrong, but it can't tell us much either. A person unwilling to commit to any interpretation never gets the benefit of discovering that an interpretation failed. An institution that leaves every question permanently open eventually becomes incapable of deciding anything.

Claims are necessary.

Commitments are necessary.

Some closures are necessary.

The trick is making enough of them to act without pretending we've thereby exhausted what can be learned.

That produces a different picture of simplicity.

A good explanation isn't necessarily small.

It's economical.

The distinctions it carries are doing work.

The assumptions expose themselves to evidence.

The abstractions help us encounter the object rather than insulating us from it.

When new information requires another distinction, we add one.

When an old distinction stops doing useful work, we can ask whether it still belongs.

Occam's Razor doesn't guarantee that we'll get those decisions right.

It can't tell us in advance which complexity will prove necessary.

It gives us a pressure in the other direction.

Before adding another layer, another category, another entity, another explanation for the explanation, ask what becomes possible because it's there.

If the answer is clear, keep it.

If reality later demands more, add more.

And if something has been sitting in the model for years without changing what we can observe, predict, distinguish or do, perhaps it's time to ask why we're still carrying it.

The razor doesn't make reality simpler.

It makes us account for the complexity we introduce while trying to understand it.