Our evidence, and its limits.

🧭Novices learn more from explicit teaching than from unguided discovery.

For learners without substantial prior knowledge in a domain, fully guided instruction — clear explanation, worked examples, step-by-step models — generally produces more learning than minimal guidance or unassisted discovery: a meta-analysis of 580 comparisons favoured explicit instruction over unassisted discovery under most conditions (d ≈ 0.38). The same meta-analysis found that guided discovery — with feedback, worked examples, scaffolding or prompted explanation — outperformed other forms of instruction (d ≈ 0.30), so the finding is against leaving novices unguided, not against every kind of inquiry. As expertise grows, guidance that helped a novice can become redundant or even harmful (the expertise-reversal effect), which is an argument for fading support, not for withholding it.

Where you can see it: Every lesson opens with a modelled worked example before any independent task, and guided practice offers hints on demand with worked solutions. Support is withdrawn ACROSS THE PHASES of a lesson — modelled, then guided with hints available, then independent with none — and within guided practice the learner decides how much help to take.

Where it is contested: Kirschner, Sweller & Clark is a genuinely contested paper — constructivist researchers argue it characterises inquiry teaching unfairly — and there is counter-evidence we rely on you knowing about. Alfieri and colleagues found guided discovery outperformed other forms of instruction, explicit teaching among them. A meta-analysis of 53 studies found that letting learners attempt a problem BEFORE instruction — productive failure — beat instruction first (g ≈ 0.36), though instruction first did better for second- to fifth-graders and for domain-general skills. Our lessons model first. We think that is defensible for novices, and for younger learners especially; it is a design choice, not a settled verdict, and for older learners the productive-failure evidence cuts against it.

🧠Working memory is the bottleneck, so presentation matters.

Working memory holds only a few elements at once. Conventional problem solving by means–ends analysis spends much of that capacity on reaching the goal, leaving little for learning how the problem works — the finding cognitive load theory began with. Instruction that splits attention across sources a learner must integrate in their head — a diagram here, the text that explains it over there — spends capacity on the format rather than the content: the split-attention effect, which Mayer's multimedia research carries as spatial contiguity. Managing extraneous load is one of the better-replicated findings in instructional design.

Where you can see it: Phases are short and single-purpose; diagrams sit with the text they explain rather than in a separate panel; a lesson never asks a learner to hold an instruction in their head while doing something else.

🤝Teach in the gap between what a child can do alone and with help.

Vygotsky's Zone of Proximal Development frames learning as social and assisted: the productive target is what a learner can do with support, not what they can already do unaided. Assistance is then withdrawn as competence grows.

Where you can see it: Progressive Diagnostics give an overall working level and a status for each strand, with what to do next; the learning page puts units in a strand that needs work first, with the reason shown, and a retake a term later is compared strand by strand. The included Learning Start consultation exists because the 'more knowledgeable other' cannot be a piece of software.

Where it is contested: ZPD is a powerful framing device rather than a precisely measurable quantity. We use it to justify assisted practice and diagnostic placement — not to claim we can measure the width of any child's zone.

🔁Retrieving beats re-reading, and spacing beats cramming.

Practice testing and distributed practice are the two techniques rated of highest utility in a large review of learning strategies — well-evidenced, generalisable across ages, materials and outcome measures. Re-reading and highlighting, the two things students actually do most, rate low.

Where you can see it: Checks for understanding are retrieval events rather than comprehension re-reads, and a missed item is re-asked once, reshuffled, after the practice phase — a second retrieval attempt, not a re-read. Scored warm-up retrieval from earlier lessons opens the capstone lesson of every unit, Prep to Year 12 — four hundred and eighty-six units, three items each. Across the whole program every wrong answer a learner can choose, 19,068 of them, carries the label of the misconception it was written to catch, so a miss is recorded as what was misunderstood rather than only that something was; where the same belief is tested in a later year it carries the same label, so a record shows a misconception recurring rather than several unrelated slips. That was retrieval at one interval. A check question from a finished lesson is now brought back on its own expanding schedule — a day later, then three, seven, fourteen and thirty days if it holds, and back to a day if it does not — six items at most in a day, the ones whose distractor named a misconception first. It sits at the top of the learner's own page, is skippable, gates nothing, and reports to nobody. Whether that changes what any individual child retains is not something we have measured, and we claim nothing about it.

✍️Producing an answer teaches more than reading one, and explaining a step teaches more than watching it.

Information a learner generates is remembered better than the same information read: a meta-analysis of 86 studies found a generation effect of about d = 0.40. Prompting learners to explain material to themselves improves learning across ages and subjects — g = 0.55 over 69 effect sizes from 64 reports. That figure does not control for time. Self-explaining usually takes longer, and in the five comparisons where both groups spent similar time the effect was g = 0.41, against 0.72 where self-explanation took longer (the difference was not statistically significant). The authors conclude it is time-efficient but warn its efficiency may be exaggerated, because most studies did not control or report time on task.

Where you can see it: A guided problem asks for the learner's own answer before the worked one opens, then for their own judgement of how it compared; hints stay available throughout, and 'I'm stuck' opens the method rather than leaving a child with nothing. Before each next step of a worked example the learner may guess it, and at the end is asked to explain one step in their own words. The longest reading stretch of every lesson closes with a pause to recall its big ideas from memory before checking them. Reflect prompts are answered rather than read, and hands-on activities record that they were done and what was noticed. In a first set of lessons whose content is a distinction — conductor or insulator, sine or tangent — the learner sorts new cases into it and reads why for any they misplaced; the first attempt is the one recorded. Those sorts are being written lesson by lesson and are not yet in most of the program. Prep to Year 2 say their answers to their helper and are never asked to type.

Where it is contested: The generation effect is measured mostly in laboratory studies with adults and word material; self-explanation is the better-evidenced of the two for school learning. Whether these prompts change what any individual child retains is not something we have measured, and we claim nothing about it.

✅Check understanding while it is forming, and feed back on the task, not the child.

Formative assessment — eliciting evidence of learning while it is still forming, and acting on it — is among the better-supported school interventions, though how large its effect is remains disputed (see below). Feedback itself is not uniformly good: a meta-analysis of 607 effect sizes found an average gain (d = 0.41), yet over a third of feedback interventions made performance worse, and effects fell as feedback drew attention away from the task and toward the self. A later review reaches the same order: feedback about the task and how to do it helps, while feedback directed at the person, praise included, rarely helps learning.

Where you can see it: Every lesson states its learning intention and success criteria up front, gives immediate feedback on checks, and closes with learner self-assessment against those same criteria. Feedback is always about the work.

Where it is contested: How much formative assessment helps is genuinely contested. The large figures often quoted for it, 0.4 to 0.7, do not survive scrutiny: Bennett found the commonly made quantitative claims trace to untraceable, flawed, dated or unpublished sources, and Kingston and Nash, able to compute effect sizes from only 13 of more than 300 studies they reviewed, found a weighted mean of about 0.2. We rely on formative assessment as sound practice with a modest and variable effect, not as a large one, and no source here compares feedback during learning with marking after it.

🔤Capable readers do four different jobs, and all four must be taught.

The Four Resources model is a conceptual framework, not an empirical finding. It describes reading as four families of practice a reader draws on together — code breaker, text participant, text user, text analyst — and argues that a capable reader needs all four, so a literacy programme should leave none of them untaught.

Where you can see it: English units and reading diagnostics deliberately cover all four roles, and each is named so a parent can see which job is being taught.

Where it is contested: Freebody and Luke offer a way of naming what a reading programme has to cover; their paper does not test whether teaching all four roles produces better readers than teaching fewer, and we cite it for no such result. We use it the way we use UDL — as a design discipline that stops a role being left out — not as evidence that our English units work. It also says nothing about how decoding should be taught, which is a separate body of evidence that this entry does not rest on.

🌱Concrete before abstract — task design introduces an idea through something a child can handle before its rule.

Inhelder and Piaget's account of the move from concrete to formal reasoning — from reasoning about things that can be handled to reasoning about hypotheses, variables and propositions — is where the developmental sequencing of the bands comes from. The STRONG version of stage theory — discrete stages arriving at fixed ages, with a child in one stage across every domain — has not held up: children reason at different levels in different domains, training and context shift performance, and the ages are far more variable than the theory proposed. Some of the strongest evidence against it shows children from 16 months to four years old inferring causal structure and forming abstract theories — reasoning Piaget placed much later. So the case for concrete before abstract inside a lesson does not rest on stage theory. It rests on instructional evidence: a systematic review of concreteness fading in mathematics and science recommends beginning with concrete materials and fading explicitly and gradually to the abstract, because the concrete grounds what a symbol means and the fading strips away the detail that would not transfer. Manipulate a thing before you name the rule; act out a scenario before you write the equation.

Where you can see it: The four developmental bands set what a lesson may assume: concrete materials and story in the early band, logical reasoning about things a child can manipulate in the middle, proportional reasoning and variables next, and hypothetical reasoning with abstract notation last. It decides the ORDER ideas are introduced in, never what any individual child is capable of.

Where it is contested: We use the sequence as a design discipline, not as a claim about any child. A band is a property of a lesson, not a diagnosis of a learner: nothing in the product reads a band off a child, and a learner working above or below their year's band is the ordinary case rather than an exception. The strong stage theory is not supported and we do not rely on it. Concreteness fading is reviewed in mathematics and science; applying the same order across every subject is our design default, and goes beyond what that review covers.

👀Learning a family watched happen can be recorded as evidence of it.

Registering authorities already say so. The NSW guidelines for home schooling registration — issued by NESA until 5 May 2025, and by the NSW Department of Education since — tell parents they can integrate informal learning and daily life as opportunities for learning, offering shopping lists against the English syllabus and measuring ingredients against Mathematics as examples — so the question is not whether daily life can evidence a syllabus, but whether a family has any way to record it. Teacher observation is likewise a long-established form of assessment in Australian practice rather than a concession: a discussion paper published by Queensland's then curriculum authority sets out its place in student assessment, and the Early Years Learning Framework embeds observation-based assessment as ordinary method. The recognised documentation forms are running records, anecdotal notes, frequency counts, work samples and learning stories. Observation is also the only instrument that reaches a great deal of real learning at all — a child who will not sit for a set task is not thereby not learning, and the task's absence measures the task.

Where you can see it: A family writes down what they saw, in their own words, and those words are kept verbatim — including in the record an authority reads. The product offers candidate content descriptions by matching the family's words against the outcomes for that year, and shows which words caused each guess. Nothing is filed until the family accepts it, and a notice that matches no outcome at all is still kept, because learning that fits no code is still learning.

Where it is contested: Observation is weaker evidence than a marked task and our record says which is which, every time, rather than blending them into one total. The learning-stories form specifically has never been psychometrically validated or tested beyond case studies, so we treat a notice as a contemporaneous record of something a supervisor saw — not as a measurement. The matching is a suggestion engine and cannot establish alignment; only the family can, because only the family was there.

🎛️Design for variability from the start, rather than retrofitting access.

Universal Design for Learning proposes multiple means of engagement, representation, and action & expression as a design discipline. It is a well-developed framework with real practical value; the evidence base for UDL as a whole intervention is still thinner than its adoption.

Where you can see it: Read-aloud on every phase, adjustable text size, reading-support font, high contrast and calm motion are in the platform itself, not an accessibility settings afterthought. Choice tasks are multiple means of expression against identical success criteria.

Where it is contested: We use UDL as a design discipline, not as an evidence claim. Reviews of UDL implementation studies repeatedly note weak study designs. It shapes how we build; it is not something we cite as proof that we work.

🌟Genuine choice raises engagement — that is why choice tasks exist.

Self-determination theory identifies autonomy, competence and relatedness as the conditions for intrinsic motivation. A meta-analysis of 41 studies found that providing choice did enhance intrinsic motivation, effort, perceived competence and task performance — so the basic effect is real and it is measured, not just theorised.

Where you can see it: Choice tasks are assessed against identical success criteria, so no option is the soft one. A learner's chosen interests only mark which options play to their strengths; every option stays on offer, in the same order, and interest never changes the level, the pathway or the expectation — that comes from the diagnostics alone.

Where it is contested: This is the claim on this page with the most awkward evidence, so read it carefully. Patall and colleagues found the effect of choice was STRONGER for instructionally irrelevant choices than for instructionally relevant ones — and our choice tasks are squarely the instructionally relevant kind. The honest reading is that offering a choice of how to demonstrate learning is likely to produce a smaller motivational benefit than, say, letting a child choose the colour of their pen. We keep choice tasks because a modest engagement gain is still a gain, because they serve Universal Design for Learning's multiple means of expression regardless, and because equal-criteria options cost the learner nothing. We do not claim they are a large effect.

🌍A task set in a child's own interest holds attention and is remembered better.

A 2024 meta-analysis pooling 34 publications on problems re-written into learners' own interests found effects on interest (g ≈ .55), on retention (g ≈ .48) and on transfer (g ≈ .36), with retention effects larger in quasi-experimental than in experimental designs. The foundational study taught fourth- and fifth-graders arithmetic and found that contextualising, personalising and offering choices raised motivation, depth of engagement and the amount learned in a fixed time — but its choices concerned instructionally incidental features, and its personalisation put personal details into otherwise identical material, so it shows surface personalisation helping, not deep. A later study in an algebra tutor found personalised problems raised triggered situational interest, which predicted accuracy and learning efficiency inside the tutor. Personalisation also had a total effect on classroom exam performance, but the study did not show that gain running through interest.

Where you can see it: A learner picks one of fifteen worlds to learn in. Where a lesson has a version written in that world, that is the version they see — the same outcome codes, the same checks, the same right answers, the same misconception labels, the same success criteria. Only the surface text changes, and the substitution engine refuses to reach anything else. A world is a doorway, never a room: level and pathway come from the diagnostics alone. Writing those versions is unit-by-unit work that has started and is nowhere near finished: today ten of the fifteen worlds have a lesson written in them, in Years 2 to 7 only. Everywhere else — every other world, and every year below 2 or above 7 — a learner sees the lesson as written. That is why the sentence above is conditional rather than a promise: it says where a lesson has a version, not that every world has one.

Where it is contested: Nearly all of this evidence is mathematics word problems in short sessions. No study tests a themed full curriculum over a year, and none tests this product. We claim what the literature shows about interest and retention; we do not claim themed delivery raises achievement here, and the transfer figure rests on six effect sizes.

🧱Our themes carry the task. Decoration is banned, and a guard enforces it.

Interesting-but-irrelevant material — the 'seductive detail' — reliably makes learning worse: a meta-analysis of 68 effects from 58 studies found an average harm of about g = −0.33. Separately, the benefit of personalising a problem to an interest depends on depth: a surface re-skin and a problem whose structure is carried by the interest are different interventions with different beneficiaries, and deeper personalisation improved affect for everyone.

Where you can see it: A themed variant must change the data of the task, not decorate it. The build fails on any variant whose text merely wraps the base unchanged, differs from it by fewer than three words, or touches only titles and labels — `node scripts/curriculum-audit.mjs themes`, run on every push. If the theme could be deleted without changing the task, it is deleted before it ships.

Where it is contested: The guard catches the mechanical forms of decoration; it cannot judge whether a re-written problem is a good problem. That remains an author's and a reviewer's job, and a themed variant that passes the guard has cleared a floor, not a ceiling. Nor do we claim the harm of decoration falls hardest on beginners: the seductive-details meta-analysis does not report prior knowledge as a moderator, and names it only as a question for further research.