# Don't Tell Me What I Already Know
# Book: Language
# Chapter URL: https://programmer.ie/books/language/17-chapter/
# Prose fingerprint (sha256, normalized): 631d0f89472fcc3a52f2422e38d6f8a140844c73a268104e3b667a4739cf0006
# Snapshot generated at build time by Hugo. This is the complete rendered prose
# of the chapter; it is not a transcript of the source file.

Iris asks about the constructed consensus-protocol talk. (A consensus protocol is a method that lets computers agree on a single value.) The system has some evidence that she is an expert. She has worked on consensus systems for two years and has given talks on the subject. So it removes the speaker’s definition of partial synchrony from her digest, on the reasonable assumption that she already knows it.
The speaker, however, is using a particular timing assumption for the demonstration. The assumption is a delay bound that the protocol can use, not merely a bound that exists but is not known in advance. Iris knows the distinction between these two readings. What she does not know is which reading this speaker chose. She sees the later performance claim without its local definition, and she concludes that the talk makes a broader guarantee than it actually establishes. The system has omitted the one piece she needed to interpret the evidence.
This is a constructed failure, not a measured one. The classic literature describes more than one partial-synchrony model, so a known bound is not a universal definition of the term. The timing condition alone also does not tell us whether safety fails when delays exceed it; that would require the protocol’s proof. The narrower and more defensible lesson is enough: expertise in a subject does not mean familiarity with a particular source’s assumptions.
The term was familiar. Its intended use in this source had not yet been established for the reader. The previous chapter fixed the task and showed why different tasks need different evidence. This chapter holds both the source and the task constant. Two people can still need different amounts of explanation because their prior knowledge differs. So this chapter asks a narrower question:
Given the same source and the same task, what would actually add something for this particular person?
The question is deliberately narrow. The proposal is not to build a portrait of Iris. It is to make one testable decision: what can be shortened without making the remaining material harder to understand or less faithful to the source.
A model of what you know, and nothing else Answering the question may require a model of the reader’s knowledge for this task. This is a significant step beyond adjusting a view to the current question, so the model should be kept as small as the decision allows. Three different kinds of model are easily confused:
MODEL OF WHAT YOU KNOW ≠ MODEL OF WHAT YOU LIKE ≠ MODEL OF WHO YOU ARE This chapter needs only the first, and even that only in a limited form. The system does not need to infer personality, identity, demographic traits, politics, preferences or learning style. It needs evidence about whether this person can use a particular concept, in the sense and context the current source requires.
Being experienced in consensus research does not answer whether Iris has seen this speaker’s definition. Disliking diagrams does not answer whether she understands a term either. The model should record what evidence exists for a specific capability. It should not generalise from expertise to a kind of person.
The transformation proposed here is:
same source same task different prior knowledge ↓ different useful representation Call this adaptation to prior knowledge rather than broad personalisation. The change is local. The system expands, collapses or shortens material based on a narrow estimate of knowledge, and it keeps whatever the selected claim requires. The change is not an instruction to optimise engagement, build a permanent profile or infer new interests.
A narrow knowledge estimate is still information about a person. It must not be retained or shared merely because it would be useful to the renderer.
Novelty is relational too The previous chapter defined relevance in relation to a unit (a passage or claim in the source) and a declared purpose. Novelty needs a further relation, between three things: the unit’s specific claim or sense, the reader’s evidenced knowledge, and the current context. Novelty is not a free-standing score stored on a document.
ESTIMATED NOVELTY(unit, evidenced knowledge, sense, context) Consider the talk’s definition of a fault as a crash failure. Iris may already know that category of fault and still need to know that this talk excludes other fault models. Tom may be familiar with the historical terminology but need the local definition in order to compare papers. Mina may need both an explanation and an example.
To isolate the role of prior knowledge, imagine all three trying to interpret the same claim for the same declared purpose. Their needs may still differ, but none of them can safely lose the clause that limits the claim. A familiar concept and a source-specific restriction are not interchangeable.
Three things are easy to run together, and they vary independently:
new in the source ≠ needed for the task ≠ new to the person A unit may introduce nothing new to an expert and still be necessary for interpreting the next claim. Conversely, a novel aside may be irrelevant to the declared task. Novelty does not override relevance, and neither overrides a preservation obligation. The variables are kept distinct because a single importance score generated by a model could hide which one justified an omission.
What the system may say it knows The system does not observe knowledge directly. It observes evidence: answers, demonstrations, corrections, self-reports and interactions, collected within whatever scope the person has authorised. Each observation supports only a limited inference.
The useful question is therefore not “does Iris know consensus?” It is “what evidence suggests she can interpret this concept in the sense the present source requires?” Framing the question this way keeps the model tied to a specific decision that can be audited.
evidence about knowledge ≠ knowledge itself A simple known/not-known flag hides too much uncertainty. For a prototype, use three operational estimates. These are working categories for deciding what to show, not diagnoses of a person’s mind:
KNOWN sufficient recent evidence for this concept and sense, under a stated policy PARTIAL evidence of some understanding, or evidence that does not cover this use UNKNOWN insufficient or conflicting evidence to justify a knowledge-based omission These grades are not a calibrated probability and should not be presented as one. Their meaning depends on a declared threshold and on the kind of evidence available. A recent mistake can conflict with an earlier demonstration, and an apparent success can depend on hints.
When evidence conflicts, the safest result for omission is UNKNOWN. The conflict should be recorded, not silently replaced by the more recent or more convenient observation. A changed source definition is a scope mismatch, not a fall in the person’s expertise.
Several smaller distinctions stop the model from claiming more than it has:
not observed ≠ not known mentioned before ≠ mastered familiar term ≠ known meaning The first guards against inferring ignorance from missing records. Iris’s experience may be invisible to the system. The second guards against treating exposure as mastery. The third is the opening failure. There is a fourth boundary as well: knowing a standard term is not the same as knowing the particular convention a document assigns to it. A useful knowledge record therefore identifies the sense and the source context, not only the concept label.
What evidence would justify even a narrow estimate? A person’s explicit statement that they know a concept is strong evidence about how they want to be treated. By itself, it is not a test of mastery. A correct explanation or a successful application is more direct evidence of capability for that task, although it may have been assisted or may depend on context. A corrected error can reveal both understanding and a prior gap. A question asked earlier may show curiosity rather than ignorance. Repeated exposure, clicks, watched videos or the mere appearance of a term in conversation show little about retention. No universal ordering of these kinds of evidence exists apart from the construct being assessed, the conditions of the observation and the person’s consent.
Age matters too. But a fixed decay rate for all topics is a policy choice, not a psychological finding. The model should keep the observation type, scope, source and date, so the system can explain why it believes something. When an old observation is the only one available, it should prefer uncertainty.
The minimum useful record already holds more than a concept and a grade. It might contain the concept, its intended sense, the task or source scope, the evidence type, the evidence location (if retention is authorised), the observation date, any uncertainty or conflict, and the person’s corrections. A grade without the evidence and sense that produced it becomes a disguised dossier entry.
A record should stay inspectable and reversible. In plain terms, it should be able to say: the system believes you can explain quorum in this setting, based on this dated answer; you can correct that estimate or decline to use it.
This is still personal information, so data minimisation is an obligation now. It is not something the book can defer until its broader privacy discussion. A session-only estimate is a reasonable default for the experiment. Any persistence across sessions, access by another system or transfer between projects needs an explicit purpose and control. Deleting a knowledge estimate should not require the reader to prove that the estimate was wrong. Later chapters examine the broader personal mediator. This chapter does not assume one already exists.
Known is not safe to omit The opening failure supplies a rule: never omit a source-specific definition, exception, assumption or prerequisite merely because the reader appears to know the underlying topic. Familiarity may justify shortening an explanation of the general concept. It does not establish that the reader knows what this author chose to mean by the term.
Consider three units in a chain, where each depends on the one before:
Constructed dependency example. Knowledge evidence cannot waive a source-preservation obligation. Open full-size diagram. Suppose the system has evidence that the reader understands A and B in their usual senses. But C depends on B as defined in this source, and that version of B has a narrower condition. Removing B because its label is familiar breaks the dependency needed to interpret C. The fact that the reader once answered a general question about B correctly does not repair the gap.
Omission therefore needs more than a knowledge estimate. It needs three things: a source-specific dependency check, the reader’s permitted degree of compression, and a decision about whether the claim stays recoverable if the explanation is hidden. A useful conservative rule is:
evidence supports this sense as known AND omission does not remove an obligation or necessary dependency AND the reader permits this degree of compression ↓ CANDIDATE FOR OMISSION (not automatically safe) Presentation should be more graduated than the choice between omitting and showing. Strong evidence of knowledge may justify a shorter reminder when no required dependency is at risk. Partial or stale evidence favours a collapsed explanation, with its existence and provenance visible. Unknown or conflicting evidence favours keeping the explanation accessible without assuming ignorance.
A source-specific qualifier needed for the displayed conclusion belongs with that conclusion. It should not sit behind an undisclosed click. The decision has to pass the preservation obligations from Part I after the novelty policy has acted.
known ≠ safe to omit There are two different errors, and their costs depend on the task. Showing a definition someone already knows wastes attention and can sometimes obscure the decisive finding. Hiding a prerequisite or qualification can silently change the apparent claim. The second error is especially costly when the reader is making a consequential decision. But there is no universal numerical exchange rate between the two. A system should report unnecessary repetition and harmful omissions separately, with stronger protection for critical dependencies. When evidence is weak, shortening the material or leaving it available is safer than deleting it silently.
What this is not A few boundaries keep the narrow model narrow. They also separate observing behaviour from inferring a reader’s understanding.
I know this ≠ I like this I have seen this ≠ I know this I often open this ≠ I need more of this I rarely open this ≠ I already understand it Interest and knowledge are separate variables, and neither can be read directly from clicks. A novice may open five articles without understanding them. An expert may ignore a familiar field because the current task lies elsewhere. An engagement recommender optimises for behaviour such as opening or watching, not for whether a reader can reconstruct a particular claim. That makes click history a particularly weak basis for removing definitions. The proposed renderer should not import a recommendation profile and relabel it as knowledge.
There is relevant research, but it supports only a limited inference. Kalyuga, Ayres, Chandler and Sweller’s review of the expertise reversal effect (2003) reports that instructional support helpful to novices can be less useful, or even detrimental, to experts under the conditions studied. That gives a reason to test adapting explanation to demonstrated expertise. It does not show that every expert should receive less information, nor that a model can identify what this particular reader knows.
Knowledge-tracing research, including Corbett and colleagues’ work on student models (1994), likewise shows how a system can maintain estimates from task observations under stated modelling assumptions. An estimate that predicts responses is still not direct observation of knowledge. Neither body of literature validates the chapter’s proposed omission policy.
How we would test a knowledge-aware digest The experiment should hold the source, declared task and available attention budget constant while varying the knowledge evidence and the presentation policy.
Prepare a set of constructed and real, permissioned technical sources, with definitions, dependencies and qualifications checked independently. Include traps like the consensus talk: a familiar term used in a different sense, a prerequisite far from the claim, a source revision that changes an assumption, and a novel passage that is irrelevant to the task. Reviewers should identify the preservation obligations and acceptable answers without seeing the system’s outputs. Disagreements become part of the evaluation record rather than being flattened into a single authoritative answer.
Compare at least four conditions:
an unadapted but well-designed digest; a simple reader-controlled expand/collapse view; an adaptation based on broad expertise labels or self-report; the proposed evidence-scoped, sense-aware policy with a dependency gate. Keep the model, source access and generation budget comparable across conditions. Record whether the knowledge observations were authorised, current and genuinely available at the time of the decision. Test across readers with different demonstrated knowledge, and include an option to decline adaptation. Do not manufacture a knowledge record from a reader’s browsing history after the fact.
The primary safety measures are harmful omissions (missing source-specific qualifications), wrong answers to task questions and false confidence. Measure unnecessary repetition, task completion time, effort and reader preference separately. Check whether readers can recover a hidden prerequisite, and whether corrections to the knowledge record change later decisions as intended.
Evaluate the accuracy and calibration of the model’s knowledge states only against an independently collected, appropriately scoped set of tasks. Do not treat a grade as “known” merely because the system assigned it. Report results by task and by expertise, including cases where a simpler collapse control matches or beats the knowledge model.
This experiment is proposed. The reference example below does not run it.
Running it Running it. The existing program holds dated, hand-authored evidence about two readers. It treats a knowledge fact as scoped to a concept and a sense, not as a global claim about a person’s expertise. The grade (known, partial or unknown) is computed from the supplied evidence and a declared age policy. The absence of an observation yields UNKNOWN, not a judgement of ignorance. Being shown material is not evidence of mastering it, and a later mistake can conflict with an earlier demonstration.
In the constructed program, Iris’s record treats one sense of partial synchrony as known: a delay bound exists without its value being known in advance. The speaker’s local assumption is stronger, so the definition is kept. This demonstrates the intended sense-mismatch rule. It does not demonstrate that a model reads technical literature correctly.
Her quorum definition is marked known and has no declared dependency in the selected output, so the constructed policy may omit it. Mina merely opened a page mentioning quorum, so her definition stays. Tests also retain a known prerequisite needed by a displayed claim, and they collapse a partially known explanation rather than deleting it. The person can inspect the record, correct it and ask the example to forget it. These are behaviours implemented in the example. They are not evidence of how readers actually use them.
What it does not show: that the evidence accurately measures anyone’s knowledge, that the sense labels were extracted reliably, or that the omission rule improves comprehension. The knowledge states are supplied by hand, not learned from real readers. The example has no durable cross-session store, and its decay periods are declared policy rather than empirical estimates. Its dependency gate protects only the dependencies it was given. It does not test privacy controls, reader outcomes or the comparison proposed above. Storage and access questions come later in the book, but no such capability should be inferred from this scripted example.
Where this leaves us Personal novelty depends on a unit’s claim or sense, the reader’s evidenced knowledge and the context. That makes it a useful hypothesis about what may add value. It is not permission to erase whatever looks familiar.
The knowledge estimate is neither a direct observation nor a proxy for identity or interest. It can justify a shorter explanation when the evidence warrants it. But source-specific meanings, qualifications and necessary dependencies must survive, and the reader must keep control over whether adaptation is used. The system may repeat something an expert already knows. The more serious failure is to hide the very condition needed to understand the rest.
Even a correct knowledge estimate is not enough to decide how information should be presented. Two people with the same knowledge may need different forms. One may be scanning between meetings while the other studies carefully, or one may be preparing to act while the other is trying to remember. These are features of the present task and circumstances, not proof of a stable personal identity. The book returns to situational adaptation later, when it can also address what the system is permitted to infer.
Before that, Part II must confront a simpler change. So far, the person has asked the question. A person’s stated interests can persist between questions, and the next useful thing may be something they have not yet thought to request.
References Dwork, C., Lynch, N. & Stockmeyer, L. (1988). Consensus in the presence of partial synchrony. Journal of the ACM, 35(2), 288–323. DOI 10.1145/42282.42283. (Distinguishes alternative timing assumptions; not evidence of the fictional talk’s actual guarantees.) Kalyuga, S., Ayres, P., Chandler, P. & Sweller, J. (2003). The expertise reversal effect. Educational Psychologist, 38(1), 23–31. DOI 10.1207/S15326985EP3801_4. (Background on expertise-conditioned instruction, not this system’s performance.) Corbett, A.T., Anderson, J.R., Carver, V.H. & Brancolini, S.A. (1994). Individual differences and predictive validity in student modeling. In A. Ram & K. Eiselt (eds.), Proceedings of the Sixteenth Annual Conference of the Cognitive Science Society, pp. 213–218. (Knowledge-tracing precedent; not a validation of this chapter’s policy.)