If publishers can invest in the Knowledge Space rather than attempting to manufacture every possible answer, an immediate question follows:
Don’t organizations already possess enormous quantities of machine-readable data?
They do.
Governments publish datasets. Manufacturers maintain product systems. Retailers operate catalogs and inventory feeds. Financial institutions maintain pricing databases. Universities maintain program information. Healthcare organizations maintain benefits, provider, formulary, eligibility, and enrollment data.
Much of that information is already structured.
But structure and knowledge are not the same thing.
A machine-readable format can make information easier to store, exchange, retrieve, and process. It does not necessarily preserve enough meaning for that information to remain useful when removed from its original system and applied to a human question.
That distinction became increasingly important in our production work.
The problem was not simply getting data into a form a machine could read.
The problem was determining how much of its meaning needed to travel with it.
3.1 Machine-Readable Is Not Machine-Usable
Consider a value:
$15.26
There is nothing difficult about making that value machine-readable. It can be represented as a number in a database, spreadsheet, API response, JSON object, or HTML document.
But what does it mean?
Add a label:
Premium: $15.26
We know more.
We still do not know enough.
Premium for what? One insurance plan? An average across several plans? Medicare Advantage? Prescription drug coverage? Monthly or annual? For which geography? During which year? For which population? According to which source?
The value is readable.
It is also precise.
It may even be perfectly accurate.
None of those properties establish that it can be used correctly in response to a particular question.
This is a common characteristic of operational data. Source systems are designed around the needs of the organizations that create them. A government dataset may organize information according to contracts, reporting periods, geographic codes, product identifiers, or regulatory categories. A manufacturer may organize products according to models, SKUs, configurations, revisions, and distribution channels.
Those structures can be perfectly appropriate for their intended purposes.
Human information needs are organized differently.
A Medicare beneficiary does not ask for records matching the appropriate contract, plan, segment, county, and plan-year fields across several federal datasets.
The beneficiary asks which plans are available where they live.
Somebody has to bridge the distance between those two representations of the same domain.
Making the source data easier to retrieve does not eliminate that requirement. An API can transport a value perfectly while leaving its meaning dependent on knowledge that exists somewhere else.
The distinction is simple:
Structure makes data processable. Meaning makes it usable.
3.2 Meaning Accumulates
Return to the same value:
$15.26
Add a label:
Premium: $15.26
Now establish what kind of premium:
Average Medicare Advantage premium: $15.26
Add geography:
Average Medicare Advantage premium in Putnam County, Ohio: $15.26
Add time:
Average Medicare Advantage premium in Putnam County, Ohio, for the 2026 plan year: $15.26
Define the population being measured and identify the authoritative source and version from which the value was calculated.
We now have something substantially different from the number with which we began.
The number did not change.
Its meaning did.
The value now participates in an assertion about a particular subject, within a particular geography and time, describing a defined population and grounded in an identifiable source.
Those distinctions determine what can safely be done with it.
Remove the geography and an otherwise correct value may be applied to the wrong place.
Remove the time period and a correct historical value may be presented as current.
Lose the population definition and an aggregate may appear to describe something it does not.
Lose the source and the assertion becomes harder to evaluate, reproduce, or update.
The problem is not that the number becomes false when one of those elements disappears.
The problem is that its meaning becomes incomplete.
That distinction matters whenever information travels beyond the environment in which its meaning was originally understood.
Inside an organization, software, business rules, documentation, and subject-matter expertise may supply that meaning automatically. A person familiar with the data knows what a field represents. An application knows which records can be compared. A business rule knows which configuration applies. A reporting process knows which version of the source is current.
An external machine does not necessarily inherit those assumptions merely because it can retrieve the value.
If the publisher expects the information to remain useful outside its original system, enough of that meaning must travel with it.
3.3 The Anatomy of Machine-Usable Knowledge
Our production work repeatedly exposed several kinds of meaning that needed to remain associated with the underlying information.
The first is the most obvious: facts.
Facts are the assertions from which answers can be constructed. A camera weighs 658 grams. A plan has a particular premium. Twenty-nine PPO plans are available. A university program requires thirty credit hours.
But every assertion immediately raises another question:
What does it describe?
That is identity.
A correct fact attached to the wrong entity is no longer a correct answer. Product families contain variants. Organizations have subsidiaries. Plans can exist within contracts and segments. Similar names can refer to different things. Identity establishes the subject of the assertion at the level necessary for its use.
Then comes context.
Where does the assertion apply? When? For whom? Under what conditions?
Geography, time, jurisdiction, eligibility, product configuration, population, and other conditions can determine whether an otherwise valid fact belongs in the current information environment.
Knowledge also contains relationships.
A product belongs to a product family. A plan is offered by an organization. A segment belongs to a parent plan. A camera uses a particular lens mount. A county belongs to a state.
Those relationships are not merely navigational conveniences. They are part of what the publisher knows about how things in the domain connect.
Assertions also have provenance.
A value may come directly from an authoritative source, be normalized from one, or be calculated from several source records. The distinction matters. Provenance establishes where an assertion came from and what supports its authority.
Finally, some information cannot safely be applied simply because it has been found.
Several values may all be true. Several variants may all be valid. The available information may still be insufficient to determine which one applies to the question being asked.
That introduces resolution.
For now, we need only recognize the requirement. We will examine its consequences later.
Taken together, the anatomy is straightforward:
Facts — What is asserted?
Identity — What does the assertion describe?
Context — Where, when, for whom, and under what conditions does it apply?
Relationships — How does the subject connect to other things necessary to interpret it?
Provenance — Where did the assertion come from, and what establishes its authority?
Resolution — Is enough known to determine which applicable assertion can safely be used?
These elements do not transform information through some formal threshold at which data suddenly becomes knowledge.
They identify something more practical.
They are dimensions of meaning that may need to survive when information leaves the environment in which that meaning was previously supplied implicitly.
A useful representation of the progression is therefore:
FACT
+ IDENTITY
+ CONTEXT
+ RELATIONSHIPS
+ PROVENANCE
+ RESOLUTION
─────────────────
MACHINE-USABLE KNOWLEDGENot every fact requires every dimension to be stated independently. Some meaning may already be established by the surrounding structure. Some domains are more conditional than others. Some assertions are straightforward enough that little additional context is required.
The objective is not maximal metadata.
It is sufficient meaning.
3.4 The Publisher Already Knows More
Much of the knowledge described above already exists.
The organization may simply not think of it as something it publishes.
People inside the organization know that one identifier represents a parent entity and another represents a child. They know that a price applies only in a particular market. They know that two differently named fields refer to the same concept. They know which source is authoritative. They know when a value expires. They know that one configuration is incompatible with another.
Applications know some of these things too.
Business rules encode them.
Databases imply them through relationships.
Editorial policies preserve them.
Data pipelines enforce them.
Subject-matter experts apply them without necessarily writing them down each time.
The knowledge exists operationally.
What may be missing is a public representation in which enough of that meaning remains attached to the information when an external machine encounters it.
This leads to a principle that will recur throughout this book:
Do not make the machine infer what the publisher already knows.
That is not an argument against inference.
Generative systems are useful precisely because they can interpret language, connect information, identify patterns, perform calculations, and synthesize conclusions that a publisher never explicitly authored.
The publisher does not need to perform all of that work in advance.
But there is little advantage in forcing every consuming system to independently reconstruct distinctions the authoritative publisher already possesses and can communicate.
The publisher represents what it knows.
The machine reasons from what is available.
That division of responsibility brings us to another problem.
Facts do not become coherent knowledge merely because identity, context, relationships, provenance, and resolution conditions exist somewhere on the same website. Meaning depends on the appropriate elements remaining associated with one another.
Some facts belong together.
Others do not.
Some context governs an entire collection of assertions. Other context applies to only one. A relationship may connect two otherwise separate subjects. Provenance may establish the authority of one assertion or thousands.
If knowledge depends on preserving those associations, then we need to determine where one coherent body of knowledge begins and ends.
What belongs together?
A page cannot answer that question merely by existing.
Neither can a paragraph.
Neither can a database row.
We need a semantic boundary.
That is where the knowledge object begins.