How to Write a Scene-Based Question: Design Notes on the mgpages Tests

Why we never ask “I tend to plan ahead”

The familiar personality-test item asks how strongly you agree with a sentence like “I tend to plan ahead.” With items like that, the person you would like to be usually answers first. If planning sounds good, you agree one notch more. I chose to hand you a scene instead of a sentence, and to hide the scoring behind the choices inside it.

The first question in Isekai Reincarnation Appraisal asks how you ended up under the truck. The Heartwarming Human Shield who dove in to save a child, the Smartphone Zombie who was reading a web novel and never saw the truck coming, the Mysterious Truck Shopper who researched which truck looked the comfiest in advance. The surface is a joke, but every choice spreads exactly 3 points across six indicators. The human shield is 2 instinct and 1 guard, the smartphone zombie is 3 instinct, and the truck shopper is 2 strategy and 1 vanguard.

The Apocalypse Survival Test goes a step further. Here a single choice raises some abilities while cutting others. Walk to the window at 4 a.m. without turning on the light, and your awareness and composure go up while execution drops by a point. Every question is built from trade-offs like that, so among all 65,536 possible answer combinations, none raises all six abilities at once. Making the model answer impossible to find is the core device of a scene-based question.

Six questions measure, the last one chooses

Three of the tests, Isekai, Rofan and Villain, share the same skeleton. Six three-choice questions accumulate points on six indicators, and the indicators pair up into three axes. For Isekai those are party versus lone, instinct versus strategy, and vanguard versus guard. Each axis collapses into a single difference between its two indicators, and the sign of that difference decides one letter. Three letters give 2×2×2, which is eight types.

Say that in Isekai you pick the human shield, “Ha! That would be me,” grabbing the elf’s hand during the duel, tossing one of the brawlers into the magic circle to test it, the cringe-worthy telepathy and the defensive barrier. You end up with 6 party, 0 lone, 4 instinct, 0 strategy, 1 vanguard and 7 guard. The three axes lean toward party, instinct and guard, and your world code becomes BIG.

The seventh question carries no points. It asks for your last words as you lie dying after the plan has failed, and the answer picks one of three roles inside the same world. If that same person chooses to confess who they really are, the result is The Monster Keeper of the Refuge. Eight types times three variations gives 24 results. In Rofan the same job is done by the thought that crosses your mind as you die exactly as the original novel said, and in Villain by the question of where you will take your place.

Ties happen more often than you would think

Deciding letters by sign leaves one gap. When the two indicators are exactly equal, the difference is zero and there is no sign. It sounds rare, but when you actually count, it happens on each axis with a probability of 7 to 14 percent. So every axis has a default side to fall to on a tie.

Those defaults were not picked by taste. With six three-choice questions there are only 3 to the power of 6, or 729, possible answer combinations. I applied each of the eight candidate tie rules to all 729 combinations and kept the one that spread the eight types most evenly. In Villain, only one of the eight candidates kept every type between 10 and 15 percent.

To check the spread, a simulation feeds 50,000 sets of random answers through the algorithm and confirms that each type lands between 10 and 15 percent, against an expected 12.5. In Villain, though, one type sat at 10.15 percent, right against the floor, so the check occasionally failed purely on sampling luck without a single line of code changing. Now a test that computes all 729 combinations decides pass or fail, and the random simulation stays on only as a report for human eyes.

All of this assumes that everyone answers at random. Real answers cluster one way or another, so what the check guarantees stops at this: no type is structurally unreachable, and no type runs away with the results.

Two tests that speak in percentages

Apocalypse, which shows a survival probability, and My Caveman DNA, which shows a caveman DNA percentage, are built differently. Eight four-choice questions award points to six abilities and to two hidden penalties that never appear on screen, recklessness and rigidity. In Apocalypse, when someone in your group clutches their arm and says they were not bitten, choosing “Believe them, act normal” adds 3 to social and 2 to composure, takes 3 away from judgment, and piles 3 onto recklessness.

Here the scores are multiplied rather than added. A weighted average of the abilities is multiplied in turn by fatal flaws, by pair strengths that switch on only when two abilities are high together, by a final clutch bonus when composure and judgment are both excellent, and by the hidden penalties. If one ability hits bottom, the whole score is cut no matter how good the rest is. Because it is multiplication, the result screen can separately compute and show what your number would have been without that weakness.

A sum of eight answers naturally bunches up in the middle. So the score is stretched onto 1 to 99 percent by blending its rank among all combinations with the original scale. Apocalypse gives the rank a weight of 0.85, Caveman 0.5. Rank alone would let a single answer swing the number too far and would erase the rarity of the top and bottom grades. That number is then divided into eight grades, from S+ down to F.

The variation is not asked; it is calculated from the abilities. Three pairs of abilities are compared, but their averages differ, and a raw comparison in Apocalypse skewed the split to 41, 36 and 23 percent. So each pair is standardized against its own distribution before the largest is chosen, and a tie follows a fixed order. Caveman still shows a final role question on screen, but the variation is now set by the eight answers before it, not by that answer.

An even spread is not the same as a correct one

The biggest lesson of this work was that a model can pass its distribution check and still be wrong. For a while, in Isekai, Rofan and Villain, the subtraction that computes each axis ran the wrong way round. Someone who loves a crowd was scored as a loner. Flip every sign and the type distribution stays just as even, so the simulation caught nothing. In Isekai the weights had been written backwards too, and the two errors cancelled out so neatly that everything looked right.

Since then, the tests check meaning. They confirm that choices talking about “together” or “companions” never score for lone, and that choices talking about “analysis” or “observation” never score for instinct. In Apocalypse they check that the choice where your body locks up does not raise composure, and that choices which leap in without checking earn recklessness. Caveman’s verification script goes as far as confirming that an answer set built to maximize one ability really does make that ability the highest.

Those checks turned up a real problem. When Apocalypse’s ability weights were carried over to Caveman unchanged, choosing the physical, stick-with-the-group answers actually lowered your caveman score. The choices that raised those abilities were cutting the others by more. I rebalanced the weights so that all six abilities push the final number upward.

Result writing, languages, and what these tests are not

Results are written as stories, not short type blurbs. The Monster Keeper of the Refuge is a scene in which a forty-five-year-old druid, trying to shield his creatures from the imperial army, blurts out a lie he can never take back, and it runs to about a thousand characters in Korean. Most of the 24 Isekai results are about that long, and each ends with advice. Villain’s Sulking Prankster shows how the trait plays out in everyday life, then suggests handing over one sentence instead of a trap: “What you said earlier — it actually stung a little.”

Isekai is published in Korean, Japanese and Chinese; the other four in Korean, English and Japanese. Caveman’s Japanese translation is still being filled in, so any gaps show in Korean. There is no sign-up, and the result screen can be turned into an image right in your browser, then saved and shared.

The limits belong here too. The weights are values I set by reading each scene, not item coefficients validated against a sample of responses. The distribution checks assume random answers, and Rofan’s survival percentages and the Villain Index are fixed values attached to each result. That is why these tests are entertainment, not psychological assessment. What I can promise is that the same answers always produce the same result, and that I can explain which rules produced it.