Every explanation encodes a set of assumptions about what its reader already knows. When those assumptions are wrong, the explanation does not fail loudly. It fails silently, and the learner discovers the failure only under examination conditions.
An explanation is always written from somewhere. It chooses a starting point, an analogy, a level of formality, and a body of prior ideas it treats as settled. For readers whose existing knowledge matches those choices, it works, and it works so cleanly that its author has little reason to revise it. For the rest, nothing announces the breakdown. There is the sensation of following along, and then, some days later, an inability to reproduce any part of it.
Teachers working one to one see this within seconds, and their response is almost never to repeat themselves. They change register: from the symbolic to the pictorial, from the general rule to the single case that motivated the rule, from the formal statement to the reason anyone bothered to formulate it. Benjamin Bloom's 1984 comparison of tutored and conventionally taught pupils placed the advantage of individual tuition near two standard deviations. The magnitude has been argued over since. The mechanism has not: a tutor can detect the moment an explanation stops landing and can produce a different one on the spot.
Gick and Holyoak's 1983 experiments on analogical transfer remain the cleanest demonstration. Participants given a single analogous story rarely applied its structure to a new problem. Participants given two stories with different surface details, and prompted to compare them, induced the underlying schema and transferred it far more often. One instance teaches the instance. A second instance, deliberately unlike the first, is what makes the shared structure visible.
Shaaron Ainsworth's DeFT framework, published in 2006, sets out why several representations of one idea do work that no single representation can. They carry complementary information and support complementary processes; a familiar representation constrains how an unfamiliar one is read; and holding two together supports abstraction and extension. The same framework names the cost. Translating between representations is itself effortful, and representations that are merely adjacent rather than genuinely related impose load and return nothing. Multiplicity is not the virtue. Relatable multiplicity is.
Rittle-Johnson and Star's 2007 work on comparing solution methods sharpens the point further. Learners who studied two methods presented side by side gained more in flexibility and conceptual understanding than learners who studied the same two methods in sequence. What mattered was the act of comparison, not the number of exposures. Mayer's programme of multimedia research, resting on Paivio's dual-coding account, points the same way: words with a coordinated diagram outperform words alone, provided the two are integrated rather than merely co-present.
The claim needs separating from a popular idea that the evidence does not support. Pashler, McDaniel, Rohrer and Bjork's 2008 review found that the meshing hypothesis, the proposition that instruction works best when matched to a learner's preferred modality, lacked the experimental support such a widely held belief would require. The argument here is the opposite in structure. It does not hold that a given learner has one correct register and should be routed to it. It holds that each register carries genuinely different content, and that most learners need more than one.
A diagram makes structure salient. Algebra makes generality salient. A worked case makes procedure salient. Plain prose makes purpose salient, and purpose is the thing most often missing when a candidate can execute a method but cannot recognise where it applies. These are not four translations of one message. They are four partial views, and the idea itself is what survives the comparison.
Sweller and Cooper's 1985 finding that studying worked examples beats solving equivalent problems, for learners new to a domain, is among the most replicated results in instructional research. Chi and colleagues showed in 1989 that the benefit depends on what the learner does with the example: those who explained the steps to themselves gained substantially more than those who read them. Renkl's later work on faded examples showed how to provoke that behaviour deliberately, by removing steps one at a time until the learner is solving unaided.
Kalyuga's expertise reversal effect completes the picture and complicates it. The support that helps a novice actively impedes a learner who has moved past it, because processing redundant guidance competes with the work of solving. A serious system therefore cannot simply hold several explanations of one idea. It has to move a learner across them as competence grows, from the fully worked to the partially faded to the bare problem, and to withdraw scaffolding rather than accumulate it.
Competitive entrance examinations concentrate their demands far more than their syllabuses suggest. Across 8,267 authentic JEE Main questions from 2002 to 2025, 117 topics carry enough questions to describe a stable pattern: 48 in Physics, 41 in Chemistry, 28 in Mathematics. Matrices and determinants alone account for 8.8 per cent of JEE Main Mathematics since 2016, 275 questions across the archive. Kinematics accounts for 6.6 per cent of recent Physics, 180 questions. Coordination compounds account for 8.7 per cent of recent Chemistry, 171 questions. When a small set of ideas is asked repeatedly in unfamiliar disguises, the return on a durable, transferable representation of any one of them is collected many times over.
Negative marking raises the cost of the fragile alternative. Break-even guessing accuracy, the penalty divided by the sum of reward and penalty, is 20 per cent for NEET and JEE Main and 25 per cent for UPSC Prelims. A blind pick from four options is exactly neutral in the UPSC case and mildly positive in the other two. Candidates who half-understand an idea do not guess blindly, though. They guess along the line of the one explanation they were given, and well-constructed distractors are built precisely from the simplifications that a single explanation tends to leave behind.
One explanation per idea is not a pedagogical position. It is an economic constraint inherited from print and from timetabled teaching, where a page and an hour were genuinely scarce. Software removes that scarcity but does not automatically replace the editorial discipline that scarcity enforced. Several explanations that restate one another in slightly altered wording are worse than one, because they consume attention and add nothing the research says learners can use.
The demanding version is to treat each register as carrying its own content, to hold the registers against a shared anchor so that comparison is possible rather than merely permitted, and to sequence support so it recedes as a learner becomes able to work without it. That is harder to build than a single authoritative account. It is also the only arrangement consistent with what decades of research on representation and transfer actually found.
The full set, with progress tracking and five agent perspectives per question, is in the JupiteX app — browse the exam catalogue or browse the Learn library.