01. The Question
Knowledge has to be written down in a form a machine can reason with.
Knowledge that used to live in people, in precedent and in habit now has to be written down so that a model can apply the policy, an agent can make the call, and a system can explain itself to an auditor.
The usual answers treat this as one problem with one solution: fine-tune a model on the documents, retrieve the right passage, add a knowledge graph. This thesis argues it is not one problem.
What form must a domain's knowledge take for machine reasoning over it to be trustworthy, and what decides that form?
02. Two Properties
How soon the answer arrives, and how often the decision is made.
Ground truth is the answer that would settle whether a decision was right. Verifiability is how, and how soon, that answer becomes available. It comes in three grades, and they are different conditions, not points on one line.
Checkable
The answer exists and can be checked at the time of the decision.
Whether a transaction meets a published eligibility rule.
Delayed
The answer arrives later, months or years, and is mixed with other causes, so isolating it takes deliberate design.
Whether a discount policy improved retention.
Unrevealed
The answer never fully arrives, or arrives in a form that cannot separate a good decision from a lucky one.
Whether a succession plan chose the right successor.
Most frameworks merge the second and third grades into “uncertain.” This thesis separates them, because they call for different representations: a delayed answer can still be tested; an unrevealed one cannot.
Frequency is how often the class of decision is made: daily, weekly, quarterly, or once in a decade.
Both are properties of a decision class, not of an industry. Within tax, applying a rule to a transaction is checkable; choosing a planning strategy is delayed; judging how a regulator will read an ambiguous provision is close to unrevealed.
03. The Spectrum
Each grade asks for a different written form.
Checkable
The answer is available now
- Written down
- Rules compiled into formal logic
- Each decision returns
- A proof trace
- The form is for
- Correctness
Delayed
The answer arrives late, and confounded
- Written down
- Typed beliefs that can be tested
- Each decision returns
- An update to those beliefs
- The form is for
- Learning rate
Unrevealed
The answer never fully arrives
- Written down
- Records of how judgments were reached
- Each decision returns
- An inspectable account
- The form is for
- Consistency and transferability
Each column makes a different promise. A proof promises the answer is right. A belief promises the organization will learn faster than it would otherwise. A reasoning record promises nothing about correctness, only that the judgment can be inspected, challenged and repeated by someone else.
A model producing fluent answers with no written form underneath fits none of the columns, which is why it can appear to work in all of them.
04. Using It
Three questions place a decision.
- Q1
Can the answer be checked when you decide?
Yes →Checkable - Q2
If not, will it arrive later, untangled enough to test?
Yes →Delayed - Q3
If neither, the answer never fully arrives.
Then →Unrevealed
Three invented decisions
Does this expense claim meet the travel policy?
Which inbound leads get a same-day call?
Which finalist becomes head of design?
A rough placement is the point; grades are judgments. Finer placement is not covered here.
05. Four Claims
What the thesis asserts.
- 01
Representation follows verifiability.
The form knowledge must take is set by how verifiable the domain's ground truth is: rules where the answer can be checked now, testable beliefs where it arrives late and confounded, records of reasoning where it never fully arrives.
- 02
Value follows frequency.
A formalization is worth most where the decision is made often and its outcome is slow to reveal whether it was made well: often enough for a written form to compound, slowly enough that nothing will learn the domain unaided.
- 03
Mismatch fails in both directions.
Too strong a form, pointed at a domain with nothing to check against, produces machinery without feedback. Too weak a form, pointed at a domain that has a right answer, produces fluency without correctness.
- 04
The representation varies; the need for structure does not.
At every point on the spectrum, a formalization has to say what must never be violated, how decisions follow from that, and how they are applied to a specific case. Only the form each of those takes changes.
06. What Would Prove It Wrong
The thesis fails if any of these holds.
F1One representation suffices.
A single approach, such as a model that retrieves sources and cites them, reaches the error rates auditors and regulators accept in checkable domains, and also produces judgments as consistent and transferable as written reasoning records in unrevealed ones.
F2Outcome learning converges where it shouldn't.
Systems trained only on outcome data reliably outperform experienced operators on frequent, delayed decisions, with no explicit beliefs underneath.
F3Mismatch is cheap.
Organizations using belief-style formalizations pass audits in checkable domains without difficulty, or formal verification produces measurable value in unrevealed ones.
Stated in October 2026 so they can be checked against what follows. F2 is the one most likely to be tested soonest.
07. Scope and Limits
Positions move. Better instrumentation can turn a delayed domain into a checkable one; a regulation can become unrevealed at its ambiguous edges. A formalization has to be re-placed when the verifiability of its decision class changes.
This is about form, not authorship. The thesis says what a formalization should look like. It says nothing about who should write it, who should check it, or whose interests it should serve. Those are separate questions with separate arguments.
Grades are judgments. Placing a decision class on the spectrum is itself an estimate, made roughly rather than precisely. The value is in distinguishing the three grades, not in scoring them to a decimal.
08. Grounded In
None of the ingredients is new. The question asked of them is.
Polanyi, The Tacit Dimension (1966)
“We can know more than we can tell.” This thesis asks what partial articulation is still possible at each grade of verifiability, and in what form.
Autor, Levy & Murnane (2003)
Codifiability decides what gets automated. Here codification is not binary: its form is set by verifiability, and non-routine tasks can still be formalized, as beliefs or records rather than rules.
Hogarth — kind and wicked learning environments (2001; with Lejarraga & Soyer, 2015)
In wicked environments, experience teaches the wrong lessons. Learning has to move out of individual experience and into a written form the organization can test or inspect.
Kahneman & Klein (2009)
Skilled intuition needs an environment regular enough to be learned and adequate opportunity to learn it. Those conditions track verifiability and frequency, applied here to machines and institutions rather than individuals.
Snowden & Boone, Cynefin (2007)
Cynefin addresses how to act. This thesis addresses what persists after the decision: the written form that makes the next one better, checkable or consistent.
The kind–wicked distinction tells you whether experience will teach you. The spectrum tells you what to write down when it won't.
09. Downstream
Other work on this site starts from here.
10. How to Cite
https://icebergtribe.com/thinking/the-knowledge-formalization-spectrum
References
- Polanyi, M. (1966). The Tacit Dimension.
- Autor, D., Levy, F. & Murnane, R. (2003). The Skill Content of Recent Technological Change. Quarterly Journal of Economics, 118(4).
- Hogarth, R. (2001). Educating Intuition.
- Hogarth, R., Lejarraga, T. & Soyer, E. (2015). The Two Settings of Kind and Wicked Learning Environments. Current Directions in Psychological Science, 24(5).
- Kahneman, D. & Klein, G. (2009). Conditions for Intuitive Expertise: A Failure to Disagree. American Psychologist, 64(6).
- Snowden, D. & Boone, M. (2007). A Leader's Framework for Decision Making. Harvard Business Review.