One Size Fits Whom? Testing Universal Teaching Policies Against Learner Variation

AA programme document says that all teaching on a course must use method X. The method has good evidence behind it: trials, a meta-analysis, a favourable review from a body that people trust. The average effect is positive, and the policy is adopted. Should anyone worry about the learner for whom method X does not work?

The worry has an intuitive form. Learners differ in what they know, how they process information, what languages they use, what their bodies and attention allow, and what is happening in their lives this week. A single prescribed method cannot be equally good for all of them, so any such policy will let someone down. That is the starting hypothesis here, and this page treats it as a claim to test rather than a position to defend. The investigation changed it, and the way it changed is the part worth following.

The page is also a worked example of the method described in Thinking in Public: Evidence Traces, Tool Calls, and Auditable Reasoning. The hypothesis is turned into proof obligations, competing hypotheses are set against it, each source is asked a specific question, trust is attached to particular contributions instead of to people, and the conclusion comes with the conditions that would reverse it. Where a move is our own interpretation rather than something a cited source claims, the text says so.the method is not the page, but its engine

The Opening Problem FoundationalKnowledge that endures for decades — core principles

ImagineImagine five people on the programme with the prescribed method.

• One arrives without the prior knowledge the method's first session assumes.
• One cannot access the representation the method relies on, because the diagrams or the audio are not usable for them.
• One is experienced, and finds the prescribed sequence slows them down for no gain.
• One is working in a second language, or has a disability, anxiety, a caring role or a bad month that changes what they can take in.
• One belongs to a group that was barely present in the studies that supported the method.

A strong average effect is compatible with all five of these stories. An average is a summary across people; it does not say that each person reached an adequate outcome. That gives the central question: does learner variation make every fixed teaching policy liable to leave someone behind?

The Initial Hypothesis FoundationalKnowledge that endures for decades — core principles

StatedStated in its strongest plausible form:

H1  Any policy prescribing a single teaching approach will
    inevitably under-serve some learners, because learner
    populations are heterogeneous.

The words any, single, inevitably and some make this far stronger than the familiar observation that learners differ. "Any" covers every policy. "Single" means a substantially invariant treatment. "Inevitably" claims necessity, not risk. "Some" can be satisfied by one learner. A claim built from those four words is easy to feel and hard to establish, which is what makes it a good subject for an investigation.

Turning the Hypothesis into Proof Obligations Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

AA claim of this size has to be broken into parts that can be examined separately. Each part below is something that would have to be true for H1 to hold. Open one to see what the investigation found.

Competing Hypotheses Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

AA claim tested only against a featureless null is easy to confirm, because almost any pattern of evidence fits somewhere. The better test sets it against alternatives that can explain the same observations. Five are used here, and several can be true at once.

• H0, universal treatment. At least one substantially invariant instructional treatment can be adequate for every learner in the relevant population.
• H2, common foundation. Some practices benefit a wide range of learners, but their implementation must stay responsive.
• H3, implementation. Most apparent learner-by-method differences come from implementation quality, resources and teacher expertise, not from inherent limits of a common pedagogy.
• H4, goal-relative. A substantially common method may be adequate for some narrowly specified learning goals even if it is inadequate across education generally.
• H5, adaptable policy. A policy can prescribe responsiveness, alternatives and feedback. Prescription is then not equivalent to one-size-fits-all treatment.

HYPOTHESIS EVOLUTION

  Every prescribed method fails someone
        |
        v
  Fixed treatments create a risk of differential disadvantage
        |
        v
  Average benefit cannot establish universal adequacy
        |
        v
  Policies carry an obligation to test variation and adaptation

Starting Beliefs Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

BeforeBefore looking at evidence, the starting position was written down. These are priors in the Bayesian sense: beliefs held before the evidence, which the evidence then revises. They are given as ranges, since a single figure would claim a precision the starting judgement does not have.ranges avoid false precision

{
  "absolute_policy_hypothesis":          [0.35, 0.55],
  "invariant_treatment_risk":            [0.60, 0.78],
  "common_foundations_plus_adaptation":  [0.65, 0.82],
  "average_effect_is_insufficient":      [0.70, 0.86]
}

These are elicited starting beliefs, not measurements. Another investigator may reasonably choose different ones, and where they do, the disagreement should be kept and tested rather than averaged away. The point of writing the priors down is to make the assumptions inspectable. What matters in the end is whether reasonable priors converge once they meet the evidence, and the control further down the page lets a reader check that for themselves. A prior is not a "trust score". It is a starting position, and the evidence moves it.

How Trust Was Assigned FoundationalKnowledge that endures for decades — core principles

TheThe evidence below comes from sources of very different kinds, and some conventions about trust need setting out first, because the usual shortcut does not work.

The shortcut is a permanent score per source, something like Trust(Source) = 0.82. It fails because reliability is not a property of a person or a journal as a whole. A researcher who is excellent in computing education may know nothing useful about football forecasting. A student may be the best possible witness to their own access barrier and a poor witness to the prevalence of that barrier across a population. A trial statistician is the reverse. Neither contribution outranks the other, because they bear on different claims.trust is not a person's property, it's a relation

The situated alternative treats trust as a probability that a particular contribution is epistemically useful for a particular purpose:

\[ R_{s,c,d,k,x,t} = P\big(\text{contribution is epistemically useful} \mid s,c,d,k,x,t,\mathcal{H}\big) \]

In ordinary language the terms are these. \(s\) is the source or contributor. \(c\) is the claim being considered. \(d\) is the domain. \(k\) is the epistemic task, such as reporting direct experience, estimating a population effect, or interpreting a mechanism. \(x\) is the context, and \(t\) is the time. \(\mathcal{H}\) is the relevant history of the source in that kind of setting. The result is a prior over a specific contribution, and it can be revised as evidence arrives.

TRUST SCOPE

  Contributor
     x Domain
     x Claim
     x Epistemic task
     x Context
     x Time
     = Situated trust prior

It also helps to pull apart things that "trust" usually collapses: domain competence, competence at the particular task, sincerity, methodological adequacy, applicability to the present population, independence from other evidence, calibration, and recency. Two consequences follow. A low-applicability study is not necessarily a poor study; it may be excellent evidence about a different population. And correlated sources are not necessarily dishonest; they may simply share the same underlying studies.

Three Kinds of Difference FoundationalKnowledge that endures for decades — core principles

OneOne inference runs through most arguments in this area and is invalid: learners differ, therefore every learner needs a different instructional method. Three distinct things can differ, and only one of them supports the strong claim.

1. Baseline variation. Learners begin with different likely outcomes: \(Y_i(0) \ne Y_j(0)\). Two people may differ in what they would achieve without the intervention at all.
2. Treatment-effect variation. The intervention affects learners differently: \(\tau_i = Y_i(1) - Y_i(0)\), and \(\tau_i \ne \tau_j\). The same method helps one learner more than another.
3. Avoidable differential disadvantage. A learner would experience a materially better outcome under a feasible alternative: \(Y_i(A) < Y_i(B)\), where \(A\) is the prescribed method and \(B\) an available alternative.

The original hypothesis is fully supported only by evidence of the third kind. Showing that learners have different attainment, preferences or backgrounds (the first kind) is insufficient, and so is showing that effects vary (the second) unless a better alternative exists and could have been used. The notation is the standard potential-outcomes framework from causal inference. Its use here is our choice of framing, not something the cited studies adopt.the third kind is interaction: method x learner

Do Learners Differ in Ways That Matter? Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

TheThe same holds for prior knowledge. That learners do not start from the same place is widely accepted. It tells us that learners need different things from a given lesson. It does not show that they need different underlying methods.

Aptitude-Treatment Interactions Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

TheThe natural next question is whether learners merely perform at different levels, or whether they respond differently to different instructional treatments. Educational psychology has a name and a long history for the second possibility: aptitude-treatment interaction, or ATI. The idea is that a treatment which suits learners with high aptitude on some measure may do less for those with low aptitude, and the reverse for another treatment, so that the right choice depends on the learner.crossing lines in a plot reveal the interaction

This evidence reduces confidence in any easy claim that learner characteristics can be measured and then matched to the right method. It does not reduce confidence in learner variation itself, which is a separate matter.

A Falsification Result: Learning-Styles Matching Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

TheThe most popular argument for learner-specific teaching is that people have fixed preferred modalities (visual, auditory, kinaesthetic and similar) and learn best when teaching matches them. If this were true it would be the clearest case of a treatment-by-learner interaction. It is also a claim that can fail.

This result cuts one way and no further. A popular argument offered in support of learner-specific teaching does not survive scrutiny. The failure of that argument does not establish that invariant teaching is universally adequate, and it would be a mistake to move from "learning styles are a myth" to "so one approach suits everyone". Learner variability is a much wider category than preferred modality.null hypothesis: no interaction

Does Adaptation Work? Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

IfIf matching to fixed types fails, does adapting teaching to learners help at all? Differentiated instruction is the standard term for adjusting content, process or support to the learners in the room. A recent meta-analysis of differentiated instruction for diverse English learners makes a useful test.

A prediction interval answers a different question from a confidence interval. The confidence interval says where the average probably lies. The prediction interval says what a new study, in a new setting, might find. When it spans zero, the sensible reading of a positive average is that it is a fact about the past studies and not a promise for the next classroom. This is the page's central argument in miniature, and it arrives from the evidence rather than from the hypothesis.prediction interval covers a single new outcome

The Strongest Counterexample: Common Foundations Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

TheThe best challenge to H1 is the claim that some teaching practices are good for almost everyone. If so, a policy built on them would not be one-size-fits-all in any troubling sense. The Education Endowment Foundation's guidance on inclusive teaching is the most prominent current statement of this view.

The distinction that matters here is between two things a policy can fix.

• Policy invariance. The policy names common principles or an instructional architecture.
• Treatment invariance. Whatever the evidence about the learners, everyone receives materially the same experience.

A common framework with formative checks and adjustments may be adaptive in its operation even though it has a single name. The second kind is the one H1 is about. A policy that says "teach using explicit instruction with feedback and scaffolding, adjusting to what pupils show" is not the same thing as one that says "use this sequence, at this pace, with these materials, for everyone". Three categories of policy follow, and the hypothesis is most defensible against the third:formative checks adjust to what pupils show

1. a policy prescribing educational purposes or broad principles;
2. a policy prescribing a common but responsive framework;
3. a policy requiring a substantially invariant instructional treatment.

Designing for Variation in Advance: Universal Design for Learning Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

UniversalUniversal Design for Learning, or UDL, tries to remove the tension by building flexibility into the design before any particular learner arrives: multiple means of representation, of action and expression, and of engagement. It is a plausible attempt. Popularity is not evidence that it works, so it needs the same testing as everything else.

Do Averages Conceal Differences? Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

AnAn average treatment effect is a single number summarising how a treatment did across a population. It can be positive while some subgroups gain much more, some gain nothing, and some lose. It can also be positive with no meaningful variation at all. Which of these holds is an empirical question, and answering it needs subgroup or participant-level analysis.

Two methods points bear on how far such analyses can be trusted. Estimating how a treatment effect varies with a learner characteristic works best with participant-level data, because analyses of aggregate study-level summaries can confuse differences within a trial with differences between trials.¹¹ And simulation work calibrated to educational trials has compared 18 machine-learning methods for estimating heterogeneous effects. It found that none predicted individual effects with high accuracy, although some could identify who benefits most or least.¹² So the tools for finding who is left behind are improving, and are still blunt.

The evidence supports this update: population-average success cannot be used to infer that every relevant learner or subgroup has been adequately served. It does not support the inflated version: every intervention with a positive average effect must contain a harmed subgroup.

What Counts as Letting Someone Down? Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

IfIf "letting down" means only attainment, the hypothesis is too narrow. Success in education includes access, participation, retention, transfer, the time and effort required, cognitive load, learner agency, wellbeing, belonging, longer-term capability, and the absence of withdrawal or non-completion. A policy could raise average attainment while imposing a severe access cost on a small group. Conversely, different attainment between learners does not by itself show harm, if each did better than they would have done under a feasible alternative.

Avoidable disadvantage needs several things to be shown together:

• the policy created or materially contributed to a barrier or deficit;
• the effect was educationally meaningful;
• a feasible alternative existed;
• the alternative would probably have improved the relevant outcome;
• the alternative would not impose disproportionate harm elsewhere;
• the policy offered no reasonable route to the alternative.

This is a demanding standard, and it should be. Those six conditions are our own operationalisation of the third kind of difference above; they are not taken from the cited studies. In United Kingdom law, a close relative of this test already exists for disabled students, and the next section sets the two side by side.

Individual Learning Plans and Reasonable Adjustments Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

EverythingEverything so far has been argued at the level of the policy. In the United Kingdom there is also a standing mechanism at the level of the individual learner, and it bears directly on the third kind of difference. The Equality Act 2010 places a duty to make reasonable adjustments on education providers, and the practical record of the adjustments agreed for a particular student is usually called an Individual Learning Plan, or ILP. (The site's page on managing ILPs across large cohorts uses the term for the record of reasonable adjustments a disabled student needs, whatever an institution's own paperwork calls it.)

The statutory test is worth setting beside the one built above. Under the first requirement in section 20, where a provision, criterion or practice puts a disabled person at a substantial disadvantage compared with people who are not disabled, the duty-holder must take such steps as it is reasonable to have to take to avoid the disadvantage.¹³ A prescribed teaching method, deadline or assessment format is a provision, criterion or practice in this sense. The correspondence with the six conditions for avoidable disadvantage is close, though the mapping is ours and not the Act's wording:

Statutory test	Condition on this page
A provision, criterion or practice puts the person at a disadvantage	The policy created or materially contributed to a barrier or deficit
The disadvantage is "substantial"	The effect was educationally meaningful
Steps exist to avoid it	A feasible alternative existed, and would probably have improved the outcome
Only steps it is "reasonable" to take are required	The alternative would not impose disproportionate harm elsewhere
The duty is owed in advance, to disabled students generally (Schedule 13, paragraph 3)	The policy offered a reasonable route to the alternative

Three features of the legal framing matter for the investigation.

• The duty is anticipatory in design. For the provision of education, Schedule 13 phrases the duty in terms of disabled students generally, not only of a named individual who has asked.¹³ The High Court has spelled out what follows. The institution need not have known of the particular student's disability for the duty to arise, and the student need not have identified the adjustment; the more obvious and serious the likely disadvantage, the stronger the case that the institution should have anticipated it.¹⁴ An institution cannot wait for each case and then improvise. That is the legal counterpart of the continuing proof obligation placed on invariant treatments earlier: the policy owes thought about the barriers it is likely to create before anyone is affected.
• The duty does not reach competence standards, but it does reach the methods of assessing them. Paragraph 4 of the same Schedule excludes the application of a competence standard, defined as an academic, medical or other standard applied to determine whether someone has a particular level of competence or ability. This is the statutory form of the first of the three policy categories: what a learner must be able to do can be fixed, while how they are taught and assessed is where adjustment applies. In University of Bristol v Abrahart the University argued that requiring oral assessment of laboratory work was itself a competence standard, because explaining and defending one's work is a core skill of a professional scientist. The student had depression and social anxiety disorder, and the oral assessments put her at a substantial disadvantage. The High Court (an appeal from the County Court, decided in 2024) agreed with the County Court that the interviews and the presentation were methods of assessing knowledge and understanding and not a standard of oral communication in themselves. The decisive evidence was in the course's own documents: the learning outcome on verbal presentation was not mapped to the module, and none of the marks was awarded for the quality of oral exposition. The judge set out three questions for any such dispute: what competence is being measured, what standards are applied to decide whether it has been met, and which parts of the process are methods of assessing the standard.¹⁴ He added that a standard may not be adjustable in an individual case, but the method of assessing it may be, and that a standard that operates in an indirectly discriminatory way must still be justified as a proportionate means of a legitimate aim. For this investigation the lesson is practical. A policy that prescribes an invariant treatment can be defended by calling the treatment a standard, and whether that description holds is an evidential question about what the course actually measures.
• The scope is disability, not variation in general. A prior-knowledge gap, a language barrier or a bad month may not engage the duty at all. ILPs therefore answer one important part of the question and leave the rest to design and responsive teaching.

The result is three layers, each answering a different part of the problem. The layering is our framing.

Layer	What it does	What triggers it	Typical way it fails
Design for variation in advance (UDL and similar)	Removes barriers many learners share before anyone meets them	The designer's anticipation	Designed flexibility that no one has tested against real barriers
Responsive teaching (formative checks, scaffolding, differentiation)	Adjusts what happens next to what learners show	Evidence in the room	Effects vary widely with context, as the differentiation meta-analysis found
Individual adjustment (the ILP)	Fits provision to a specific learner's barrier	Disclosure, assessment and agreement	Late, partial or inconsistent delivery; reliance on the learner to ask

It would be convenient to conclude that ILPs close the gap, so that no policy needs further scrutiny provided it permits adjustments. The investigation does not support that, and the reasons are specific.

• They depend on disclosure and assessment. A learner with an undiagnosed or undisclosed need, or with a temporary circumstance, has no plan. The route exists for those who enter it.
• They place the burden on the learner. Someone already at a disadvantage has to ask, evidence the need and wait, although the law does not require the student to identify the adjustment or the institution to have been told first. A barrier that is repeated across many plans is better treated as a defect in the policy than as a series of separate needs.
• A plan is a record, not a delivery. A reasonable adjustment that sits in a file but is not applied by every marker or tutor has not, in the sense the Act cares about, been made. Consistency across staff is a real source of failure.
• Labels are a poor guide to need. A good plan names a barrier and an adjustment, not a type of learner. Inferring need from a category repeats, at the individual level, the error of the learning-styles argument.
• Adjustments can reduce challenge. The inclusive-teaching guidance discussed above warns that some well-meant adaptations lower expectations or make independence harder to develop. An adjustment should change access to what is being learned and leave what is asked of the learner as it was.

The plans also have a use beyond the individual case. Recurrent adjustments are data. If the same adjustment appears in many plans for a module, that is evidence of a barrier in the module's design, and the stronger response is to remove it for everyone through design in advance. This is an interpretation, not a finding from any cited study, but it joins the layers up: the individual route supplies the evidence that keeps the design layer anchored in real barriers, and the design layer reduces the volume the individual route must carry. Trust follows the same scoping as before. The court put the point in similar terms: what a disabled person says and does is evidence, although it can be reasonable to verify it or ask for expert recommendations before acting, and urgency or severity can make the account sufficient on its own. A student's account is strong evidence that a particular barrier exists, a professional assessment speaks to the nature of the need and a suitable adjustment, and the course team is the authority on what the standard actually requires. Each contribution discharges a different obligation.

Updating the Beliefs Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

AfterAfter the evidence traces, the starting ranges were revised. These are the investigator's judgements, made with the traces in view. They are not frequencies extracted from the literature, and no dataset or defined model produced them.

{
  "original_absolute_claim": {
    "posterior": [0.24, 0.43],
    "decision": "not supported in its absolute form"
  },
  "invariant_treatment_risk": {
    "posterior": [0.79, 0.90],
    "decision": "supported as a risk claim"
  },
  "common_foundations_plus_adaptation": {
    "posterior": [0.82, 0.92],
    "decision": "strongly supported"
  },
  "average_effect_is_insufficient": {
    "posterior": [0.88, 0.96],
    "decision": "strongest conclusion"
  }
}

The evidence items were not treated as independent. Reviews can contain overlapping primary studies. Institutional guidance can draw on the same syntheses. Later papers repeat earlier assumptions. A credible update has to discount for this, and citing five sources that all rest on the same three trials should move belief less than five independent ones would. Counting citations as votes is the error the discount avoids. Numbers such as these do not create objectivity: priors, likelihood judgements and trust assessments all remain contestable, and their value lies in being written down where they can be argued with.

Do Reasonable Priors Converge? Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

ExpressedExpressed as a single combined weight of evidence for each claim (a likelihood ratio, after discounting for dependence), the revisions above amount to roughly: 0.6 for the absolute claim, 2.5 for the risk claim, 2.5 for common foundations plus adaptation, and 3.5 for the insufficiency of averages. These weights are implied by the ranges; they are not additional measurements. The control below lets a reader choose a different starting belief or a different weight and see what follows.

Moving the sceptic's starting point shows the limit of the exercise: a strong prior against a claim is only partly moved by evidence of this strength. Priors within the stated range converge on the same decision. Priors far outside it may not, and the right response is to say so, then ask which further evidence would move them.

A Decision Tree Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

TheThe investigation can be compressed into a decision aid. It is an aid to deciding what to ask next, not a deterministic proof, and several of its branches end in "open obligation" for good reason.

Does the policy prescribe aspects of teaching?
|
+-- No
|     The hypothesis does not apply.
|
+-- Yes
      |
      +-- Does it prescribe broad purposes or an invariant treatment?
            |
            +-- Broad purposes or responsive principles
            |     Not necessarily one-size-fits-all.
            |
            +-- Substantially invariant treatment
                  |
                  +-- Is the population heterogeneous in ways relevant
                        to the task?
                        |
                        +-- Not demonstrated
                        |     Insufficient evidence of differential
                        |     disadvantage.
                        |
                        +-- Yes
                              |
                              +-- Can implementation respond to
                                    identified needs?
                                    |
                                    +-- Yes
                                    |     Test whether the available
                                    |     adaptation is sufficient
                                    |     (design in advance, responsive
                                    |     teaching, individual
                                    |     adjustments), and who must
                                    |     ask for it.
                                    |
                                    +-- No
                                          |
                                          +-- Is there evidence of a
                                                material barrier or
                                                heterogeneous effect?
                                                |
                                                +-- No
                                                |     Risk remains
                                                |     plausible but
                                                |     unconfirmed.
                                                |
                                                +-- Yes
                                                      |
                                                      +-- Is a feasible
                                                            better
                                                            alternative
                                                            available?
                                                            |
                                                            +-- No or unknown
                                                            |     Open proof
                                                            |     obligation.
                                                            |
                                                            +-- Yes
                                                                  Avoidable
                                                                  disadvantage
                                                                  supported.

The Assurance Case Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

AnAn assurance case is a structured argument for a top-level claim, in which each supporting claim has evidence or a status attached. The conclusion map below records what the investigation supports, what it refutes, and what it leaves unestablished. Showing the failures alongside the successes is deliberate.

TOP CLAIM
Rigid instructional prescriptions create a significant risk of
avoidable disadvantage in heterogeneous learner populations.

|-- Learners vary in educationally relevant characteristics
|     Supported
|
|-- Some instructional effects vary by learner and context
|     Supported, but difficult to estimate
|
|-- Average positive effects establish universal adequacy
|     Refuted
|
|-- Every learner difference demands another teaching method
|     Refuted
|
|-- Learning-style matching solves learner variability
|     Not supported
|
|-- Common pedagogical foundations can be broadly beneficial
|     Supported
|
|-- Common foundations remove the need for adaptation
|     Not supported
|
|-- Individual adjustment routes (ILPs) make further scrutiny
|   of a policy unnecessary
|     Not supported: they depend on disclosure and delivery
|
|-- Education providers have a standing duty to make
|   reasonable adjustments for disabled students
|     Supported (statute)
|
|-- Every policy prescribing teaching harms someone
|     Not established
|
`-- Policies should monitor access, contextual variation and
    heterogeneous effects
      Strongly supported

The Next Most Informative Investigation Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

AA further literature search would add citations without necessarily changing anything. A more informative next step is a test applied to a concrete policy, one that can fail. The worksheet below follows the steps of a policy stress test. It is meant to be reused on a real institutional policy, and the entries stay in the reader's browser.

The learner-context cases for step 6 are experimental scenarios, not fixed identities. Candidates include:

• strong domain knowledge with limited proficiency in the language of instruction;
• limited prior knowledge combined with high working-memory demand;
• sensory or motor access requirements;
• fluctuating attention caused by temporary circumstances;
• an experienced adult learner with entrenched misconceptions;
• a novice requiring substantial explicit structure;
• a learner unable to participate synchronously;
• a learner for whom the examples depend on unfamiliar cultural assumptions;
• a student with a recorded plan whose adjustments are not applied by every marker.

The same person can be in several of these at different times, which is the point of calling them cases and not types.

Where the Investigation Ended FoundationalKnowledge that endures for decades — core principles

TheThe original hypothesis did not survive unchanged. Learner diversity does not show that every educational policy prescribing a teaching approach must fail someone, and it does not support sorting learners into fixed types. Some common instructional principles appear useful across diverse groups, particularly when they contain mechanisms for checking understanding and adapting what happens next.

What learner diversity does undermine is the inference from average success to universal adequacy. Where policy converts a useful principle into an invariant treatment, it takes on a continuing proof obligation: to look for barriers, inspect variation in outcomes, test plausible counterexamples and provide routes to adaptation. A policy should not be trusted because it is described as evidence-based. Its evidential standing depends on which learners, outcomes and contexts were examined, what alternatives were considered, and whether the policy remains capable of learning from the people for whom it does not work.

This is also why trust cannot be stored as a permanent score attached to a source. Trust is a prior over a particular contribution in a particular domain, made for a particular epistemic purpose. An agent's task is not to decide who is trustworthy in general, but to keep an inspectable record of why this contribution was given this weight in this investigation, and what future evidence would change that judgement. The calculus of trust treats the related problem of how trust composes along a chain of sources.

Open Obligations and Reassessment Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

SeveralSeveral obligations remain open. The Education Endowment Foundation guide's own wording on adaptation has not been read in full (trace 5). The pooling inside the learning-styles papers and the study-level data behind the differentiation and UDL reviews were not examined (traces 3, 4 and 6). The Free School Meals figures come from abstracts and a synthesis page (trace 7). The effectiveness of particular adjustments and of ILP delivery was not examined (trace 8). Nothing here tests a concrete policy; that is the work of the stress test above.

The conclusion would change under any of the following:

• dependable evidence of invariant methods producing equivalent outcomes across relevant learner conditions;
• strong evidence that adaptation produces no meaningful benefit, or creates greater harm;
• repeated demonstrations that apparent subgroup effects disappear under adequately powered analysis;
• better measures of individual treatment effects;
• evidence that a proposed universal framework reliably accommodates relevant access and contextual variation;
• discovery that the policy under examination permits more adaptation than its wording first suggested.

Related Topics

• Thinking in Public: Evidence Traces, Tool Calls, and Auditable Reasoning — the method this page applies: hypotheses, checks, results and revisions as the record of reasoning.
• A Distributed Embedding Calculus of Trust — how trust can be represented and composed across sources.
• AI, Accessibility, and Learning — access barriers and how technology changes them.
• Do the Design Rules for Auditable Agents Survive Testing? — the same method applied to design guidance for agentic systems.
• Managing Individual Learning Plans Across Large Cohorts — the legal baseline for reasonable adjustments and what strains when plans multiply.
• Pedagogy for Networked and Agentic Learning — teaching approaches for learners who work with networks and agents.
• Evaluating Detailed Rubrics: Costs, Risks, and Whether the Precautionary Principle Actually Applies — another case of weighing a standardising instrument against what it costs the people it constrains.

References