How QuiJutsu's Practice Opponents Are Built to Challenge You Without Crushing You

← Back to blog

By Navdeep Singh Brar, founder of QuiJutsu

A practice opponent has a hard job. If it plays too well, a new player loses five games in a row and never comes back. If it plays too badly, the match is a formality and a win means nothing. The goal sits in between: a game you should usually win if you know your shows, but one you can lose if you rush or guess.

This post explains how that balance is set. It covers how your rating measures you, how a practice opponent is chosen to match it, and how that opponent decides when to answer and whether to get it right. Then it shows what happened when those opponents met real players over two days in late October, and what I changed afterwards.

Your rating measures results, not answers

Ranked uses an ELO rating. Every account starts at 1200. After each match the game works out how likely you were to win, based on the gap between your rating and your opponent's, and moves your rating by how surprising the real result was. Against an equal opponent you were expected to win half the time, so a win moves you up by half the step size and a loss moves you down by the same. Beat someone rated 400 above you and you gain almost the whole step.

The step size shrinks as you play. It is 40 points for your first 30 Ranked games, 24 until 200, and 16 after that, so a new account finds its level quickly and an established one moves slowly.

Your rating never looks at a single answer. It only sees who won. But winning in QuiJutsu comes from answering correctly and answering fast, so over enough games the rating ends up tracking both. The charts further down show this clearly: players who were matched against opponents in higher rating bands answered more text questions correctly and answered faster than players in the lower bands. Your rating is a measure of results, and results turn out to be a good measure of accuracy and speed.

Who you meet when nobody else is queueing

If no player is found within 30 seconds, a practice opponent joins. There are 20 of them, rated from 1000 to 2400. The queue picks the ones closest to your rating and chooses randomly among any that are within 50 points of the closest, so you do not always meet the same face. Their ratings never move, which keeps every band staffed no matter how many games are played.

The questions come from your own watched list, the same as any match. The difficulty follows the lower of the two levels, which in practice means almost every question a practice opponent sees is an easy one. In Ranked, a game against a practice opponent is scored as if the opponent had exactly your rating, and only half the change is applied. You cannot gain a lot by beating one that is badly placed, and you cannot lose a lot to one either.

How a practice opponent answers

Each opponent has a level from E to S, just like a player. The level sets two things: how likely it is to get a question right at each difficulty, and the window of time its answer is drawn from.

On every question the opponent rolls against its chance to be right, then picks an answer time at random from its window. Each opponent also has a personal speed setting that stretches or squeezes that window, so two opponents at the same rating can feel different. Some answer quickly and slip more often, others are slower and steadier.

The answer time matters because speed is worth points. A correct easy answer is worth up to 100: the full amount if instant, half if it lands at the last moment. A wrong answer costs 30, 40 or 50. The opponent plays by exactly these rules, scored by the same server function that scores you. It does not see your answers and does not react to your score.

What happened against real players

From 25 to 27 October, players finished 1,475 matches against practice opponents. Players won 77.6% of them. The first thing I checked was whether that number held across every rating band, or whether one band was carrying it.

It held. Players won roughly three games in four at every level. The top two bands came in a little lower, though with 54 and 17 matches they are too small to lean on. Players also improved fast: in their first game against a practice opponent they won 69% of the time, and by their fourth game that had risen above 80%.

Next, accuracy. Text questions first, then visual.

Two things stand out. First, the player line climbs with the bands, from about 67% on text below 1200 to about 79% at 1500 and up. That is the rating doing its job: the players who were meeting stronger opponents were the stronger players. Second, practice opponents were close to players on text questions, and in the top band were actually more accurate. Yet they still lost seven games in ten there.

Visual questions are a large part of the reason. Players got nine or more in ten right in every band. Practice opponents sat around two in three. When you see a character's face, you either know them instantly or you do not, and players who picked a title for their list usually do.

Speed is the other part. Players answered correctly in about two seconds, and in the upper bands closer to one and a half. Practice opponents took three to five. A correct answer at two seconds keeps 90% of its points and one at four seconds keeps 80%, so over eight questions a player who knows the answers banks more than an opponent who knows them too.

Why they felt so consistent

Players told me practice opponents felt more consistent than people, and the data agreed, though not in the way it first seems. Match to match, a practice opponent's score actually varied a little more than a player's did. What made them feel steady is that their accuracy is the same on every title. A person is brilliant on the show they rewatched last month and shaky on one they finished years ago. A practice opponent is equally good at everything, which reads as calm and machine-like, and which is also why players who chose their own list tend to beat it.

What changed on 27 October

The results were close to the goal, but the two ends of the range needed attention. The bottom band is where brand new players meet their first practice opponent, and first games were the ones players lost most often, so that is where the opponent should be gentlest. At the top, opponents were answering text questions more accurately than the players they met and losing anyway on speed, which is a strange kind of game: you are outplayed on knowledge and win on reflexes.

So text accuracy now follows the opponent's rating band directly. On easy text questions, practice opponents below 1200 now land around 55%, opponents from 1400 to 1499 around 70%, and opponents at 1500 and above around 75%. Medium and hard questions are scaled down in the same proportion, so a harder question is still harder for them. The 1200 to 1399 bands already sat where they should and are unchanged, and so are visual questions and answer speed.

I tested the change by having each opponent answer 1,200 easy questions in a sandbox before it went live. They landed within about three points of every target. The charts above are from before the change, and I will compare them with the next couple of weeks of matches before touching anything else.

The point of all this

A practice opponent is a stand-in for the person who should have been in the queue. It should be beatable, it should punish careless guessing, and it should never decide your rating for you. The numbers say it mostly does that, and the change should make the first few games kinder and the higher bands fairer.

For the rest of the matchmaking rules, see how matchmaking works. Or jump into a match and see whether a real player turns up first.