How accurate is the engine¶

You get a number, and you get the list of things the number is hiding. This page is the short version. It says how close the engine is to the real game, where it is closest, where it is worst, and what a bot author should do about that.

The short answer: good, not perfect, and measured.

How it is checked¶

Real matches are recorded. The same match is then replayed in the engine, and every unit's position and hitpoints are compared against the recording, on every tick. A tick is 50 ms of game time, so a three minute battle is 3,600 comparisons per unit.

There are 73 recorded battles behind the numbers below, all from one run of the engine at build d872d792711934c2. 67 of them play. 27 are scored from start to finish, and the other 40 up to the first card that engine could not play yet. The last 6 do not play at all.

You cannot re-run this one yourself

The recordings of real matches are private, so the accuracy measurement is not something a reader can reproduce from a clone. That is a fact about the project and we would rather say it than dress it up. The speed numbers elsewhere on this site you can re-run in a minute. This one you cannot.

The headline¶

How often the engine agrees with the recording:

how often the engine agrees	counting towers	towers left out	from the run at
a unit is within a quarter of a tile of where it really was	85.1%	56.5%	`d872d792711934c2`
a unit's hitpoints are exactly right	80.8%	79.5%	`d872d792711934c2`

Every figure in this table is from one run, at d872d792711934c2, over 73 recorded battles: 270,972 unit-ticks without towers and 799,831 with them.

The engine has changed since that run. On 2026-09-26 it switched nineteen rules to measured values, and seven more cards started to load (RoyaleSim 126992a). No re-run on the new engine is published here yet, so read every number on this page as the number for build d872d792711934c2.

The contrast between the rows is the useful part. When the spawner emission point was corrected on 2026-09-22, position went from 49.4% to 56.5% and hitpoints went from 79.4% to 79.5% without towers, and from 81.7% to 80.8% with them. Correcting where a spawner puts its units changed where units stand and not how much damage they took, in both populations, and the towers-included figure moved slightly DOWN. A change that improves the thing it targets and leaves everything else roughly where it was is the shape you want; one that moves every number at once usually means something else moved too.

The right-hand column is the one to look at. Towers do not move and there are six of them in every battle, so counting them flatters the result. A tower that sits still is not evidence that the engine walks units correctly.

"Within a quarter of a tile" is a tight bar. A tile is one square of the arena grid, the unit everything on the board is positioned in, so a quarter of one is a short distance. A unit can be in the right place to your eye and still be counted as a miss here.

One unit is close. A crowd is not¶

A Knight on its own

Within a quarter of a tile 72.9% of the time. While it walks to a tower untouched, it is on the game's own path, to a fiftieth of a tile, on 73.3% of those ticks. Close to solved.
Goblins

40.6%. Cheap units arriving three or four at once are where the engine and the game come apart.
Skeletons

46.0%. Same story. The more bodies touching each other, the worse it gets.

So the gap is not spread evenly. It is almost all in swarms. Every figure in this section is from the run at build d872d792711934c2, over 73 recorded battles.

What goes wrong first¶

The same run that produces the table above also reports what went wrong first in every battle. That is the useful table, because it says what to fix. Measured at build d872d792711934c2:

cause of the first divergence	battles	share of all the misses, in battles that went wrong this way first
when a unit dies	16	31.2%
where a spawner or a multi-unit card puts its units	10	29.4%
attack timing	10	20.5%
how units push each other apart on contact	12	18.9%

Do not compare this table with the one it replaced. Contact was 31.0% and is now 18.9%; death was 21.9% and is now 31.2%. Contact did not improve. Each battle is counted against whatever went wrong FIRST in it, so correcting the largest cause changes the scene every other cause is measured on, and a row-by-row comparison of two runs measures nothing.

19 of the 67 battles in the corpus never diverge at all, though most of those are short ones.

These numbers are from the run at build d872d792711934c2, not the target. The target is that a swarm fight does not diverge either. Expect the table to move.

What this means for your bot¶

Your bot learns the engine. It does not learn the game. Those are the same thing only where the engine is right.

This section is judgement, not measurement

Nobody has trained a bot here and then tested it against the real game. The paragraphs below are what we expect from the table above, not something anyone has measured. When somebody does run that test, this section gets replaced with what actually happened.

Where the engine is closest, a bot's habits should carry over. Pushing a single tank down a lane, choosing which lane, spending elixir at the right moment, defending a lone unit with a building: those lean on walking, routes and timing, and those are the parts measured closest to the game.

Where the engine is furthest, they may not carry over. A bot that learns exactly how a pile of Skeletons flows around a Knight is learning a crowd behaviour that is 46.0% right. A habit built on that can be worth nothing, or worse than nothing, in a real match. That 46.0% is from the run at build d872d792711934c2.

The practical version, three lines:

Before you rely on a specific interaction, look at the table. If it lives in the swarm rows, treat what your bot learns about it as a guess.
Do not tune a reward function against a single exact interaction. Reward the outcome, not the choreography.
Check again in a month. The causes above are open work, and the numbers on this page are a snapshot of build d872d792711934c2.

An honest summary¶

This engine is not as accurate as running the real game, which is true by definition. What it gives you instead is that it is fast, it runs anywhere, it needs no game files, and it tells you exactly how wrong it is and where.