Likelihood Ranking

How it works

Each round is 8 questions. A question presents a scenario - a table of student test scores, a set of distribution statements, dice/card/urn events, Poisson arrival counts, or a labelled scatter plot - and a shuffled list of options to arrange from MOST LIKELY to LEAST LIKELY. You reorder the list with up/down move buttons, then submit; the reveal shows the correct order and your pairwise agreement before the next question.

Every question has its own time limit, shown as a countdown - when it hits zero your current arrangement is submitted as-is, so a half-sorted list still earns whatever partial credit it deserves. The pool mixes hand-curated questions with procedurally generated ones regenerated on every menu visit, filterable by category and by Warm-up / Interview / Hard difficulty.

Scatter questions render a real plot with axis values and dashed threshold guide lines - they are reading-and-counting exercises over labelled points, not pixel-eyeballing.

How scoring works

Your ordering is scored on concordant pairs: of all pairs of options, what fraction did you place in the same relative order as the truth? A perfect ranking (100 percent of pairs) scores 3 points, 75 percent or better scores 2, 50 percent or better scores 1, and below that 0.

A round's maximum is therefore 3 points per question over 8 questions. Your best round total is stored locally and shown on the menu.

Rank by comparison, not by computation

The task never asks for a probability, only an order - so compute as little as possible. For most pairs, a dominance argument settles the comparison without numbers: an event that is a strict subset of another is at most as likely; a prediction that continues a student's trend beats one that breaks it; a wider tolerance band catches more probability mass than a narrower one at the same center. Sweep the options for these free comparisons first.

When two options genuinely need numbers, compute both to one significant figure and compare - the order only flips if you are wrong by more than their gap. Reserve real calculation for the one adjacent pair that remains close after the cheap passes. Ranking n options is n-1 adjacent comparisons once you have a rough order, not n full probability calculations.

Carry a small kit of distribution facts

Distribution-density and Poisson questions run on a handful of anchors. For anything bell-shaped: the mode is the most likely region, about 68 percent of mass sits within one standard deviation, 95 percent within two, and density falls off fast beyond that - so "within 1 sd of the mean" beats "between 1 and 2 sd out" beats any far-tail interval, at equal widths. An interval's probability is roughly its width times the density where it sits.

For Poisson arrivals with mean lambda, the most likely counts hug lambda (the mode is the floor of lambda), and probabilities fall off asymmetrically - slower above than a naive symmetric picture suggests for small lambda. So rank counts by distance from lambda first, and when two counts are equidistant, remember small-lambda Poisson leans toward the lower count.

For dice and cards, count outcomes over the sample space. Two-dice sums rank by number of combinations: 7 is the peak with six ways, and each step away loses one way. Card events rank by counts out of 52. These are enumeration facts, not formulas - the faster you can say "ways to make it", the faster the whole option list orders itself.

Sort like an algorithm under the clock

The interface swaps adjacent items, so plan for that: find the option you are most confident is first, move it to the top, then the most confident last, move it to the bottom, and only then arbitrate the middle. Anchoring the extremes first is both cognitively easiest (extremes are where dominance arguments are clearest) and mechanically cheapest in an adjacent-swap interface.

The scoring makes partial correctness valuable in a specific way: every correctly ordered PAIR counts, so getting the extremes right earns pair credit against every other option even if the middle is scrambled. With four options there are only six pairs - top and bottom placed correctly is already five of them. When the timer is short, extremes first is not just convenient, it is optimal.

Because a timeout submits your current arrangement, never leave the list in its shuffled state while you think. Make your cheap dominance moves immediately so the standing order always reflects your best current guess; refinement can then be interrupted by the clock without catastrophe.

A worked example

Question: two fair dice are rolled; rank these from most to least likely: (A) sum is 7, (B) sum is 10, (C) sum is 3. Count ways out of 36: A has six (1-6 through 6-1), B has three (4-6, 5-5, 6-4), C has two (1-2, 2-1). Order: A, B, C.

With three options there are three pairs, and this counting settles all of them - a perfect ranking for 3 points. Note that a scrambled middle would still have paid: placing A first alone gets two of the three pairs right, which is 67 percent - just under the 75 percent band, showing why the second anchor is worth placing too.

Common mistakes

• Computing exact probabilities for every option when dominance arguments (subset events, trend-following predictions, distance from the mode) settle most pairs for free.

• Leaving the list shuffled while thinking - a timeout submits whatever is standing, so make cheap ordering moves immediately.

• Ranking bell-curve intervals by their width alone and ignoring where they sit - a narrow interval at the mode can beat a wide one in the tail, and density falls fast past two standard deviations.

• Treating two-dice sums as equally likely instead of counting combinations - 7 has six ways, 10 has three, 3 has two.

• Agonizing over the middle of the order before anchoring the extremes, which is where most of the pairwise credit lives.

Why interviews test this

"Which of these is most likely?" is one of the most common probability screen formats precisely because it tests intuition without arithmetic cover: a candidate who truly understands distributions can order outcomes faster than one who can only compute. Ranking under a clock rewards the comparison habits - dominance, mode-distance, counting ways - that interviewers probe with follow-up questions.

The pairwise scoring mirrors how these answers are actually judged in a room: a mostly-right ordering with a defensible mistake reads far better than a refusal to commit, and knowing which comparisons you are sure of is itself the calibration being tested.