By this point, we can publish a fairly sophisticated Knowledge Object.
We can give it a coherent boundary, organize it into Knowledge Aspects, use native HTML structures that correspond to the knowledge being communicated, make selected meanings more explicit, and identify public resources through established vocabulary.
None of that establishes why an assertion should be believed.
Publishers usually address that problem through attribution. An article links to a source. A table includes a source note. A methodology page explains where the data came from. A footer identifies the organization responsible for the information.
Those practices remain useful.
But once individual assertions can be retrieved, compared, recombined, or used independently of the document in which they originally appeared, “source” becomes a more complicated idea.
The source of the underlying data may not be the source of the published assertion.
That distinction is especially important in high-stakes domains. Health, financial, legal, insurance, and other consequential information often moves through several hands before it reaches the person who relies on it. An authoritative source may provide the evidence while a publisher selects, transforms, calculates, classifies, or interprets that evidence.
If those steps change the claim being made, they are part of the knowledge too.
11.1 Who Actually Said This?
Consider two statements on a Medicare resource:
Plan A has a monthly premium of $18.
The average monthly premium among qualifying plans in this county is $15.26.
Suppose both ultimately depend on authoritative CMS data.
It would be tempting to place the same source note beneath both:
Source: Centers for Medicare & Medicaid Services
That attribution is not exactly false.
It is also not precise enough to describe what happened.
CMS may have published the record establishing that Plan A has an $18 premium. If the publisher reproduces that value faithfully, the lineage is relatively direct. CMS made the underlying assertion and the publisher represented it.
The $15.26 county average is different.
CMS may never have published that number.
Instead, CMS supplied records containing individual plan premiums. The publisher selected the records that qualified for the calculation, decided how the population should be defined, applied an aggregation method, and produced $15.26.
The source data came from CMS.
The assertion did not.
That gives us three useful distinctions.
A source assertion substantially represents something asserted by an identified source. The publisher may change formatting, expand codes into readable labels, or join source records to place the fact in context without becoming the originator of the underlying claim.
A derived assertion is produced by applying an operation to other knowledge. An average, count, percentage, comparison, or classification generated from source records may be entirely reproducible while still being an assertion created by the publisher’s methodology.
A publisher assertion originates with the publisher more directly. It may be an explanation, editorial classification, interpretation, statement of canonical identity within the publisher’s own system, or another claim for which the publisher itself accepts responsibility.
These categories are not a ranking of trustworthiness.
A publisher assertion is not inherently weaker than a source assertion. In some circumstances, the publisher is the relevant authority. A derived assertion may be mathematically straightforward and highly reliable.
The distinction concerns lineage and responsibility.
If CMS reports that a plan has an $18 premium, CMS is the authority behind the source fact.
If we calculate that the county average is $15.26, CMS supplies the evidence and we make the derived claim.
If we classify that average as comparatively low, we have introduced another publisher judgment.
All three statements may belong legitimately within the same Knowledge Object.
They should not masquerade as the same kind of knowledge.
This becomes particularly important when a machine encounters the assertions independently. A source note attached to the bottom of a page may make every number above it appear to carry the same lineage even when some values came directly from the source and others were created by the publisher.
The more consequential the information, the more that distinction matters.
11.2 Provenance Is a Chain, Not a Citation
Once we distinguish the evidence from the assertion, provenance begins to look less like a citation and more like a chain.
Consider the $15.26 county average again.
Its lineage might look something like this:
CMS source records
↓
Defined qualifying population
↓
Publisher methodology
↓
Average calculation
↓
$15.26The source is part of the provenance.
So is what happened after the source.
That matters because derivation introduces decisions.
Which records qualified for the calculation? Were multiple segments treated independently? Were zero-premium plans included? How were missing values handled? Was the result weighted or unweighted? What rounding rule was used?
Two publishers can begin with the same authoritative dataset and produce different numbers without either one fabricating anything.
The difference may be methodology.
That makes a statement such as “according to CMS” potentially misleading when attached to a derived assertion. CMS may establish every source record used in the calculation without ever having asserted the resulting average.
A more faithful description preserves both roles.
CMS supplied the source knowledge.
The publisher performed the derivation.
The same principle applies outside Medicare. A financial publisher may calculate a ratio from values reported in regulatory filings. The filings establish the inputs; the publisher produces the ratio. A healthcare publisher may classify facilities according to a methodology built from government quality measures. The government supplied the measures; the publisher created the classification.
Provenance needs enough detail to preserve that chain.
The identity of the source also matters.
“CMS” may be far too broad. A large public agency can publish many datasets, each with its own releases, reporting periods, identifiers, and update schedules. A manufacturer can revise a specification. A regulatory source can issue a new edition. A database can retain the same name while its contents change over time.
The useful question is not merely Who was the source?
It is Which source state actually supports this assertion?
Time complicates the chain further.
Suppose a Medicare enrollment dataset is released in August and contains enrollment measured in July. The publisher retrieves the file in September and updates the KNOL the same day.
Which date describes the assertion?
All of them may matter, but they mean different things.
The source publication date tells us when the source released the information. The retrieval date tells us when the publisher obtained it. The measurement period tells us when the underlying observation occurred. An effective period may tell us when the published assertion applies.
Those dates should not be collapsed into a generic “last updated” label merely because one date is easier to display.
A plan benefit published in September may become effective the following January. An enrollment figure published in August may describe July. A quality measure displayed today may summarize observations collected much earlier.
The lineage of the assertion includes those temporal distinctions where they affect its meaning.
An assertion can also stop being current without becoming false.
A 2026 Medicare premium remains a valid historical assertion about 2026 after the 2027 plan year begins. What changed is not the truth of the historical statement. What changed is the period in which that statement describes the current state.
Provenance therefore helps answer more than Where did this come from?
It helps answer What produced this assertion, and under what source and temporal conditions is the claim justified?
11.3 Provenance Belongs Where the Claim Lives
Traditional source attribution often operates at the document level.
That works well when a document has a relatively simple lineage. If one authoritative source supports the entire resource, a page-level source note may tell the reader what they need to know.
A mature KNOL may be different.
Consider a Medicare plan Knowledge Object.
Its identity and benefit information may come from one CMS dataset. Enrollment information may come from another. Performance information may come from a separate source with a different reporting period. A county average may be calculated by the publisher. An explanatory paragraph may reflect publisher-authored domain knowledge.
All of those assertions can appear on the same page.
A footer saying “Source: CMS” cannot accurately describe all of their lineages.
The problem becomes even more visible when two assertions appear beside each other:
| Assertion | Value | Provenance | Plan year |
|---|---|---|---|
| Monthly premium | $18 | Source: CMS plan data | 2026 |
| County average premium | $15.26 | Source inputs: CMS plan data Method: Publisher calculation |
2026 |
The values may appear in the same table.
Their provenance is different.
If the table is later extracted, transformed into another representation, or used as evidence in a generated answer, that distinction should not disappear merely because the two numbers once shared a visual container.
This leads to a useful rule:
Provenance should attach at the lowest level necessary to remain true.
Sometimes the entire Knowledge Object shares one lineage. Object-level provenance may be sufficient.
Sometimes a particular Knowledge Aspect comes from a distinct source. Provenance can belong at that level.
And sometimes neighboring assertions have different sources, derivations, or temporal conditions. In that case, provenance needs to travel with the individual claims.
This is an architectural decision rather than a requirement to repeat a citation after every value.
Many assertions may legitimately reference the same provenance information. The implementation can avoid unnecessary duplication.
What matters is that the provenance boundary matches the actual lineage of the knowledge.
This becomes especially important for machine consumption because extraction can change the surrounding context.
A human reading a page may see a methodology note directly above a table and understand that it governs the values below. A machine retrieving one row or one assertion may not carry that note with it unless the publishing architecture preserves the relationship.
The same problem appeared earlier with identity and context.
Meaning that exists only because nearby material happens to explain it can become fragile when the information is used independently.
Provenance is no different.
If the origin or derivation of an assertion changes what that assertion means, the lineage belongs with the knowledge at the level where the distinction remains true.
11.4 Never Assert More Than You Know
Provenance ultimately establishes a boundary on what the publisher is justified in saying.
That boundary can be easy to obscure when the published information is highly structured.
A value may have a stable identifier, explicit property, canonical datatype, precise source reference, and carefully modeled relationship to the rest of the Knowledge Object.
None of those things makes the underlying evidence stronger than it is.
Semantic precision is not epistemic certainty.
Consider three statements:
CMS reports a monthly premium of $18.
We calculate an average monthly premium of $15.26.
We classify this plan as having comparatively low cost sharing.
All three may be reasonable.
They are not the same kind of claim.
The first reports an external source assertion.
The second reports a publisher derivation.
The third introduces an evaluative judgment whose validity depends on whatever standard the publisher used to define comparatively low.
A reader should not have to reverse-engineer those differences.
Neither should a machine.
This is particularly important in YMYL publishing because the consequences of misplaced authority can be substantial.
A healthcare publisher can use authoritative government data to produce an interpretation. That does not make the interpretation a government assertion.
A financial publisher can calculate a risk measure from regulatory filings. That does not mean the regulator assigned the risk classification.
A Medicare publisher can derive a county-level average from CMS records. That does not mean CMS published the average.
The distinction does not weaken the publisher’s contribution.
It identifies it correctly.
The same discipline applies to uncertainty.
Some conclusions are estimates. Some depend on assumptions. Some reflect incomplete evidence. Some represent the best available interpretation rather than an established source fact.
Those states are part of what the publisher knows about the assertion.
They should not disappear simply because the publishing system is capable of representing the resulting value with great precision.
If the source establishes an average, the publisher should not silently turn it into a claim about every member of the population.
If the source applies to a particular period, the publisher should not present the assertion as timeless.
If a classification is editorial, it should not be disguised as an external fact.
If a derivation depends on assumptions, those assumptions belong to its lineage.
And if the available evidence does not justify a conclusion, adding more structure does not create permission to make one.
The governing principle is straightforward:
Never assert more than you know, and never attribute more to a source than the source actually establishes.
That is not merely a citation standard.
It is part of the architecture of trustworthy knowledge.