The knowledge object gives the publisher a coherent boundary around something it knows. It allows facts to remain associated with the identity, context, relationships, provenance, and other meaning necessary to use them beyond the environment in which they were created.
For a while, that seemed like most of the problem.
If we could represent the right knowledge coherently and make it available to an external machine, then the machine could retrieve what it needed and construct the answer.
Production work with Medicare plan identifiers exposed the flaw in that assumption.
The machine could retrieve the right knowledge.
Every fact could be true.
And the answer could still be wrong.
That discovery introduced a distinction that became central to the architecture developed in this book:
Retrieval establishes what knowledge is available. Resolution determines which knowledge applies.
5.1 The Problem We Encountered
Medicare Plan IDs initially looked like an unusually easy information problem.
Consider:
H5216-043
The identifier contains a Contract ID and Plan ID. A person using it is not vaguely describing an insurance product or relying on a marketing name that might refer to several things. They have supplied a structured identifier pointing toward a specific part of the Medicare plan system.
Suppose the person asks:
What is the premium for H5216-043?
The expected process seems straightforward. Identify the plan, retrieve the premium, and return the answer.
The difficulty is that Medicare plan identity can extend another level.
A parent plan may contain multiple segments. Conceptually, the structure can look like this:
H5216-043
|
+-- H5216-043-1 → $0
+-- H5216-043-2 → $18
+-- H5216-043-3 → $27Assume that all three segments are valid and that each premium comes from the appropriate authoritative source.
The parent identifier is correct. The child identifiers are correct. The premiums are correct.
The machine can retrieve all of them successfully.
It still cannot safely answer the question.
The reason is that geography determines which segment applies. H5216-043 identifies the correct plan family, but it does not provide enough information to determine which geographic segment governs the person asking the question.
The machine has three true premiums.
It does not yet know which one is the applicable premium.
That distinction changed how we understood the problem.
The missing information was not another premium. The publisher already possessed all three.
The missing information was the context required to determine which true premium applied.
Retrieval had succeeded. It had established the relevant body of knowledge.
Resolution had not.
This was not merely an unusual characteristic of Medicare data. Medicare had exposed a more general information problem.
5.2 The Camera Problem
Consider a camera retailer selling a product family called the Apex X5.
The retailer maintains authoritative specifications for two generations:
Apex X5
|
+-- X5 Mark I → Maximum video mode: 4K/60
+-- X5 Mark II → Maximum video mode: 4K/120A customer asks:
Can my Apex X5 shoot 4K at 120 frames per second?
The retailer has the information required to understand the product family. It knows that the Apex X5 exists in two generations. It knows the video capabilities of each. The specifications are accurate and current.
But the question identifies only the product family.
If the system answers yes because it retrieved the Mark II specification, it has used a true fact.
That does not make the answer correct.
The assertion that the X5 Mark II supports 4K/120 is factually valid. Whether that assertion applies to the customer’s camera depends on which generation the customer owns.
If the question does not establish the generation, the available information is insufficient to make that determination.
Three different questions are involved.
The first is whether the fact itself is true. Does the X5 Mark II actually support 4K/120? If the authoritative specification says that it does, the fact is valid.
The second is whether the fact applies. Does the customer own the Mark II? If so, the specification is applicable. If the customer owns the Mark I, the Mark II specification remains perfectly true. It simply does not describe the customer’s camera.
The third question is whether we currently know enough to distinguish between those possibilities. If the customer has said only “my Apex X5,” we do not.
These are different dimensions of correctness: fact validity, applicability, and resolution sufficiency.
The distinction matters because factual accuracy by itself does not guarantee answer accuracy.
A system might attempt to avoid the problem by saying that the Apex X5 supports video modes “up to 4K/120.” That statement may accurately describe the capabilities found somewhere within the product family, but it can still mislead the owner of a Mark I.
The problem is not the specification.
The problem is its application.
The same failure occurred in the Medicare example. All three premiums could be authentic. The system could cite authoritative data for every one of them. It could even report that premiums range from $0 to $27.
But the person asking the question does not have a premium ranging from $0 to $27. One segment applies to that person.
The range describes the candidate facts.
It does not resolve the question.
This is why fabricated information is only one kind of answer failure. A system can retrieve entirely authentic information and still select, generalize, or apply it incorrectly.
Every fact can be true.
The answer can still be wrong.
5.3 Resolution Requirements Are Knowledge
The camera retailer knows more than the two video specifications.
It also knows that generation determines which specification applies.
That distinction may already be built into the retailer’s catalog, product configurator, support system, or internal product model. A knowledgeable employee would understand it immediately.
Ask whether the Apex X5 Mark II supports 4K/120 and the answer is straightforward.
Ask whether my Apex X5 supports 4K/120 and the employee may respond with another question:
Which generation do you have?
That clarification is not evidence that the retailer lacks product knowledge.
The retailer may possess every relevant specification.
What it lacks is one piece of context from the customer that is necessary to select the applicable specification.
The Medicare publisher knows something similar. It does not merely know the premiums associated with the segments of H5216-043. It also knows that geography determines which segment applies.
In both cases, the publisher possesses two kinds of knowledge.
It knows the candidate facts.
And it knows the condition required to select among them.
That second kind of knowledge is easy to overlook because organizations frequently encode it somewhere other than the published facts themselves. It may live in application logic, database relationships, configuration rules, internal documentation, or the experience of people who work with the domain every day.
Yet the rule is part of what makes the facts usable.
If generation determines which camera specification applies, that requirement is part of what the retailer knows about the product.
If geography determines which Medicare segment applies, that requirement is part of what the publisher knows about the plan.
This led us to an important conclusion:
Resolution requirements are knowledge.
The implication is not that publishers should manufacture an answer for every possible ambiguity. It is almost the opposite.
Sometimes the publisher needs to communicate the condition under which an answer becomes possible.
For the camera customer, that condition is the generation.
For the Medicare beneficiary, it may be the county.
Once the missing discriminator is supplied, the publisher does not necessarily need new knowledge. The knowledge already present can be resolved to the applicable fact.
This expands the publisher’s responsibility beyond publishing what is true.
The publisher may also need to preserve what must be known before a truth can safely be applied.
5.4 The Boundary of Valid Assertion
Chapter 4 established one boundary around knowledge.
The Knowledge Object answers the question: What belongs together?
Resolution exposes another boundary: How far can what we know safely be applied?
The camera retailer can validly assert that the Apex X5 Mark II supports 4K/120.
It can validly assert that the Mark I supports 4K/60.
It can validly assert that both belong to the Apex X5 product family.
But if a customer has identified the camera only as “my Apex X5,” the retailer cannot yet move from those valid facts to the statement:
Your camera supports 4K/120.
Something changed between those two assertions.
The first describes what the publisher knows about the Mark II.
The second applies that knowledge to a particular customer’s camera.
That application requires another piece of information.
The same boundary appears in the Medicare example. The publisher can validly assert the premium associated with every segment of H5216-043. Without the geography required to identify the applicable segment, it cannot validly assert that one of those premiums is your premium.
This is the boundary of valid assertion.
It also reveals an important difference between two kinds of incomplete information.
Sometimes the publisher genuinely does not know. The necessary source may not exist, the available data may be incomplete, or the fact itself may never have been established.
At other times, the publisher may know everything necessary to answer correctly once another discriminator is supplied.
In the second case, the knowledge is not necessarily incomplete.
The request is.
That distinction changes what a correct response can look like.
The camera retailer may ask which generation the customer owns. The Medicare publisher may need to know the customer’s county. In another domain, the missing discriminator might be a year, jurisdiction, configuration, eligibility condition, or location.
Sometimes the correct next step is not to produce another fact.
It is to determine what is still needed before one of the known facts can safely become the answer.
A system that refuses to cross that boundary is not necessarily failing to answer. It is preserving the distinction between what is known and what can currently be concluded.
That is resolution restraint.
The concept is important because generated answers can make unsupported conclusions appear complete. Fluency does not reveal whether the necessary distinctions were actually resolved before the answer was produced.
A publisher therefore has a responsibility that extends beyond factual accuracy. It must preserve not only what is true, but the conditions governing the valid application of that truth.
Retrieval can succeed. The source can be authoritative. Every relevant fact can be correct.
And a definitive answer can still be unjustified.
Retrieval establishes what knowledge is available. Resolution determines which knowledge applies.
The boundary between those two states is the boundary of valid assertion.