Examination commentary is written one paper at a time, which is the smallest unit at which almost nothing can be established. Read as a single series of 8,267 questions, the JEE Main papers of 2002 to 2025 support a narrower set of claims than the annual verdicts, and a firmer one.
Every year, within days of the paper being written, the counting starts. Question totals per chapter are compared with last year's, differences are noted, and a verdict is issued about where the examiner is moving. The exercise is not foolish; it is underpowered. A single paper is one sample drawn from a process that has been running since 2002, and the number of questions any one area receives in it is a small integer with real variation around it.
Tversky and Kahneman described this error in 1971 as a belief in the law of small numbers: the intuition that a small sample will faithfully resemble the population it was drawn from. It does not. A count that halves from one year to the next is what ordinary variation looks like at the size of one paper, and its return to the earlier level the year after is then read as a second trend rather than as regression towards a mean that was there the whole time. Commentary is generated, preparation pivots, and nothing in the examination has actually moved.
The archive behind this essay is 8,267 authentic JEE Main questions from the papers set between 2002 and 2025. At that size the arithmetic behaves differently. Matrices and determinants account for 8.8 per cent of the mathematics asked since 2016, and 275 questions across the papers since 2002. Kinematics in one and two dimensions accounts for 6.6 per cent of recent physics, 180 questions. Coordination compounds account for 8.7 per cent of recent chemistry, 171 questions.
None of those three shares could be established from one paper. Within a single mathematics section, a share of that size is a handful of questions at most, indistinguishable from an accident of that particular morning. Sustained across dozens of sittings and hundreds of questions, the same figure becomes a description of the examination itself: about one mathematics question in eleven has come from one area, year after year, through several changes of format and of administering body. That is a structural fact, and structural facts only become visible at length.
The same window also marks its own limits. Of everything the three subject syllabuses name, 117 topics carry enough questions for a pattern to be described with any confidence: 48 in physics, 41 in chemistry, 28 in mathematics. The remainder appear, but too thinly and too irregularly for a share to carry meaning. The honest reading treats them as unmeasured rather than unimportant, since an area that has produced few questions in twenty-four years can still produce one in the next paper, and a candidate who skipped it will not be consoled by its historical rarity.
Publishing that boundary costs something, because it is an admission of how far the counting reaches. It is still the more useful statement. A frequency table that assigns a confident percentage to every line of a syllabus has stopped distinguishing between what was counted and what was filled in, and a reader has no way to tell which rows are which. The number of describable topics is itself a finding: it says where the evidence ends and judgement begins.
Content shares drift slowly. The marking scheme does not drift at all, because it is published in advance. The break-even accuracy for attempting an uncertain question is the penalty divided by the sum of reward and penalty. For JEE Main and for NEET that figure is 20 per cent. For UPSC Prelims it is 25 per cent, which is exactly the accuracy of a blind pick from four options, so guessing there is neutral in expectation by design. At 20 per cent, a blind pick from four options is mildly positive, and an aspirant who can eliminate even one option is comfortably past the line.
This is not an argument for attempting everything left blank; time spent and the variance of a single sitting both matter, and neither is captured by an expected value. It is an argument about what long observation is good for. That arithmetic has held across the entire window and is knowable before the paper is opened, whereas next year's distribution of questions is knowable by nobody. Part of what a long series does is separate the features of an examination that can be planned around from those that can only be prepared for.
Frequency data is most often misused as a shortlist, with the high-share areas worked hard and the rest abandoned. The archive does not support that trade. A share near nine per cent describes where questions have historically concentrated, not what will appear on a particular morning, and no concentration in this data is strong enough to make the remainder safe to ignore. What such figures can reasonably inform is the allocation of effort: which areas earn repeated return over months, and which need a competent single pass.
How that effort is spent is settled by evidence with no connection to any one examination. Roediger and Karpicke's 2006 experiments on retrieval practice found that testing oneself on material produces markedly better retention than rereading it, even though rereading feels more productive while it is happening. Work on distributed practice, reviewed by Cepeda and colleagues in 2006, shows a comparable advantage for spacing sessions apart rather than massing them. A long window tells learners where the examination has lived. It says nothing about how to learn, and the two questions are worth keeping apart.
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.