Why Some Knowledge Can’t Be Centralized

There is a persistent temptation to think that important knowledge becomes more useful when we gather it into one place.

Often it does.

Standards allow people who have never met to build compatible things. Procedures preserve lessons that would otherwise have to be rediscovered. Medical knowledge can move between hospitals. Engineering practices can travel between companies. A measurement made in one place can be understood somewhere else because people agreed in advance about what the units mean.

Civilization depends heavily on our ability to do this.

We observe many individual cases, find something that survives comparison and compress what we've learned into a form that can travel.

The interesting question is what happens next.

A rule, standard, policy or model is necessarily smaller than the collection of experiences from which it emerged. That's why it's useful. If a procedure required everyone applying it to personally relive every failure that contributed to its creation, we wouldn't have gained much by writing the procedure.

Compression lets learning travel farther than the people who originally did the learning.

Something is discarded in the process.

Usually that's intentional.

A building code doesn't need to preserve the personality of every engineer who contributed to it. A newspaper style guide doesn't need the complete history of every editorial argument that shaped each rule. A software interface becomes useful precisely because the person using it doesn't need to understand everything happening underneath.

Abstraction hides particulars so that attention can move somewhere else.

The danger begins when we forget that anything was hidden.

A rule that worked across a thousand previous cases encounters the thousand-and-first case, and something about this one is different. Perhaps the difference doesn't matter. Perhaps the person applying the rule simply doesn't understand why it exists. Perhaps allowing everyone to improvise would recreate exactly the failures the standard was designed to prevent.

Or perhaps the new case contains information the rule doesn't represent.

We can't know merely because somebody says, “This doesn't fit.”

The exception gets a hearing, not automatic authority.

That distinction matters because centralized knowledge and local judgment can each fail in opposite directions. A large institution can become so confident in its model that contradictory observations are treated as errors in the people reporting them. A local operator can become so confident in experience that every inconvenient standard looks like interference from people who don't understand the situation.

Both can be protecting accumulated knowledge.

Both can also be protecting themselves from correction.

What matters is whether the disagreement has somewhere to go.

Imagine a procedure developed after years of experience. Most of the time it works well, including in situations where an inexperienced person might otherwise make a predictable mistake. Someone eventually encounters a case where following the procedure appears to produce the wrong result.

One possibility is to let that person ignore the procedure whenever they feel like it.

Another is to insist that the procedure must be followed because it was written by people with greater expertise.

Neither response teaches us very much.

A better system preserves both pieces of information: this rule exists for reasons, and this case appears not to fit it.

Now we can investigate.

Maybe the exception disappears once the reason for the rule becomes clear. Maybe the procedure needs an additional condition. Maybe an unusual local circumstance requires a documented exception. Maybe the rule was once correct and the environment around it has changed enough that the original solution is now causing a different problem.

The important feature isn't whether authority began centrally or locally.

It's whether reality can still modify what the organization thinks it knows.

This is where contextual knowledge becomes particularly interesting. A farmer may notice something about a field that doesn't appear in a regional agricultural model. A craftsperson may develop a technique that isn't described in the manual. A reporter who has covered the same town for fifteen years may recognize that an ordinary-looking agenda item has unusual significance because of something that happened years earlier.

Those observations can be valuable precisely because they preserve particulars.

They can also be mistaken.

The farmer can misread a season. The craftsperson can preserve a habit long after better techniques exist. The reporter can become so familiar with an old conflict that they interpret a new disagreement through it.

Proximity creates another observer position.

It doesn't grant immunity from error.

The same is true of centralized knowledge. Aggregation can reveal patterns no individual observer could see. Standards can preserve hard-won lessons after everyone who learned them firsthand is gone. A rule developed elsewhere can protect someone locally from a failure they have never personally encountered.

Distance can carry knowledge too.

So I don't think the useful distinction is between knowledge that should be centralized and knowledge that must remain distributed. Almost any useful knowledge can be represented somewhere else. The harder question is what happens to it during representation.

What distinctions survive?

What context disappears?

What assumptions become invisible because everyone involved in creating the rule shared them?

What happens when someone downstream encounters something that doesn't fit?

Those questions change how I think about documentation too.

Documentation is sometimes criticized because the written procedure can never contain everything an experienced person knows. That's true, but it doesn't follow that the knowledge should remain in the experienced person's head. If someone has learned something important through twenty years of work, preserving even an imperfect representation of it may be enormously valuable.

The document doesn't have to become the person.

It needs to carry enough that the next person doesn't begin from zero.

Then the next person adds something.

This is how centralized knowledge can remain connected to the environments where it is used. A procedure carries previous learning outward. People apply it in new situations. Unexpected cases travel back. The procedure changes when the accumulated evidence justifies changing it.

Knowledge moves in both directions.

That sounds obvious, yet organizations routinely build the first half of the path much better than the second. They become excellent at distributing policies, standards, models and instructions while having remarkably weak mechanisms for information to travel back from the people living with the consequences.

Eventually the center knows what the procedure says more clearly than it knows what the procedure does.

That is a dangerous kind of coherence.

Everything agrees because disagreement has nowhere to land.

The solution isn't necessarily decentralization. Creating a thousand independent rule sets can destroy useful coordination while reproducing the same mistakes in a thousand places. Sometimes the center really is the best place to maintain the shared standard.

The missing piece may simply be return.

Someone needs to be able to report the exception. The exception needs enough context to remain intelligible after it travels. Someone needs authority to investigate it, and the shared model needs a mechanism for changing if the evidence warrants change.

Not every complaint changes the rule.

Not every local variation deserves preservation.

Not every standard should remain provisional forever.

A bridge shouldn't be redesigned because one person dislikes the load calculation, and a unit of measurement shouldn't change when it crosses a county line.

The point is narrower than that.

Whatever we've compressed into a rule should retain some relationship with the reality that originally taught us the rule.

This is also why inherited procedures deserve more curiosity than either obedience or contempt. Something that looks unnecessarily complicated may contain the residue of failures we never experienced. Before removing it, we should try to recover the problem it was solving.

Then reality gets another turn.

Perhaps the problem still exists and the old rule remains useful. Perhaps the problem disappeared. Perhaps the rule created new costs that now exceed its benefits. Perhaps a better solution became possible.

The fact that a rule contains accumulated learning doesn't require us to preserve it forever.

It gives us a reason to understand it before deciding.

The same principle works at larger scales. Shared knowledge allows institutions and communities to coordinate across distance and time. Local experience exposes cases the shared representation didn't anticipate. Neither needs to defeat the other.

The center can remember across places.

The edge can notice what changed here.

Then information moves.

That movement is what keeps abstraction from becoming detached from consequence and local experience from becoming trapped inside itself.

We don't need every piece of knowledge to remain where it originated. One of the great human accomplishments is our ability to carry learning beyond the people and places that produced it.

We just need to remember that carrying knowledge changes its form.

A map can travel much farther than the terrain.

That's why we make maps.

We get into trouble only when the terrain changes and there is no longer any way to update the map.