Beginner — Quoridor AI Opponent

Scout is the beginner-tier opponent, the gentlest of the six difficulty tiers on the roster. It is the right opponent for your first games after the beginner's guide, for anyone still building the habit of counting both paths every turn, and for trying an idea out without being punished for getting it slightly wrong. Current ratings for every bot live on the bots page.

What makes Scout worth a page of its own is not that it is small. It is that its weakness is manufactured. The bot does not simply think less and therefore choose worse — a separate layer takes the move the search chose and, a good fraction of the time, throws it away and plays something else. Building the weakness rather than letting it fall out of a small search makes Scout's mistakes legible, and legible mistakes are the only kind you can learn to punish.

A small budget, and a flat one

Scout's allowance works out to a couple of hundred playouts per move (engine/elo.ts:82), against 60,000 at the top of the ladder (engine/mcts/engine-v3.ts:106-111) — how the engine scores a move. A sample that size is real but shallow: enough to notice that walking forward beats walking backward and that a wall across your route is worth something, nowhere near enough to see what a wall will be worth four moves from now. How a budget of any size gets divided among the moves it has to judge is set out on Mason's page.

That allowance is also a single flat cap. The per-position envelope — where a tier is allowed to spend more thought on a harder board — does not begin until advanced (engine/bot-tuning.ts:51-53, how complexity sets the budget). The consequence is one you notice in the first minute rather than one you have to take on trust: Scout answers almost instantly, and it answers at the same speed on move four and on move forty. That constant tempo is pleasant to play against, and it is also the first honest signal that nothing about the position is changing how hard it is being looked at.

The mistakes are put in on purpose

Here is the part that is specific to the bottom of the ladder. After the search finishes and picks its move, a wrapper runs (engine/mcts/engine-v3.ts:332-341) that can replace that move with a different one. It never reaches inside the search — the search itself is untouched, and the substitution happens to its output (engine/mcts/weakening.ts:161-163). Three things can happen, in order:

1. A plain march. With a probability that is at its highest at beginner — closer to a quarter of moves than a tenth (engine/mcts/weakening.ts:41) — the chosen move is discarded and replaced with a straight step along Scout's own shortest path (engine/mcts/weakening.ts:200-204). Whatever the search wanted to do about you, it does not do; it just walks toward its goal.
2. A bounded blunder. Otherwise, with another probability that is also at its ladder maximum here, Scout plays a pawn move that is deliberately worse than the best step available — but by a controlled amount. The wrapper collects every legal step, measures how far each leaves it from its goal, and picks uniformly among those that concede more than nothing and no more than a small ceiling of tempo (engine/mcts/weakening.ts:118-132). At beginner that ceiling sits at its ladder maximum of three squares of path (engine/mcts/weakening.ts:47, 92). Scout will give back a step or three. It will not hand you the game.
3. A softened choice. If neither of the above fires, Scout does not automatically take the move its search settled on. It samples among the moves the search actually considered, weighted by how often it visited each and flattened by a temperature that is at its ladder maximum here (engine/mcts/weakening.ts:41, 141-158). In practice the search usually concentrates hard enough that this returns the searched move unchanged; where it bites is the position whose top two candidates finished close, and there it is close to a coin toss.

Every substitution is checked for legality and falls back to the searched move if it fails (engine/mcts/weakening.ts:202, 208, 214), so nothing illegal or bizarre reaches the board. The point is the shape: Scout plays like a weaker player, not like a strong player in a hurry. A fast strong engine plays consistently mediocre moves. Scout plays a mixture of sensible moves and identifiable human-looking mistakes — marching when it should be blocking, giving up a step for no reason. That is much better practice, because those are the mistakes you are learning to spot and punish.

Two things sit outside all of this, and both matter when you are planning your last walls. A settled race with no walls left on either side is solved rather than searched (how the engine works), and that answer is returned before the wrapper ever runs — so it is the one stretch of the game in which Scout's move is not being corrupted at all. And Scout does not get the wall-less shortcut that starts at advanced (Warden's page): while you still hold walls, it keeps searching and keeps being weakened right to the last move.

One tier up, Mason at intermediate runs the same wrapper with every knob dialled down.

What an error actually is

An error is not a move that felt wrong afterwards. It is a measured quantity: the difference between what your move leaves you and what the best legal alternative would have left you, in the same position, on the same turn. That is precisely the arithmetic the blunder branch performs — it scores every legal step and measures concession against the best score in that set, not against some assumption about what "one step better" ought to look like (engine/mcts/weakening.ts:119-127).

Three things follow from that definition, and all three trip up new players:

• A move that gains no ground is not automatically an error. When a wall goes down in front of you, the first step of your new best route is often sideways, and sometimes it points away from your goal entirely before it turns. It still shortens the path you actually have to walk, and if it is the best step on the board it costs you nothing. Row progress is not the measure; the length of the route left is.
• A move that gains ground can still be an error. If a better step existed and you took a lesser one, you conceded tempo even though your pawn got closer. Progress is not the standard; the best available is.
• Errors have sizes, and the sizes are the interesting part. A one-square concession and a four-square concession are not the same mistake with different luck attached; they are different mistakes. Scout makes the small kind on purpose and almost never makes the large kind, which is exactly why it is a useful sparring partner: it lets you practise noticing the small ones while there is still time to notice them.

The comparison is always against the alternatives that existed on that turn, never against how the game turned out. A move that lost you the game three turns later is not thereby an error, and a move that happened to work is not thereby correct. This is the one habit worth carrying up the whole ladder, because it is the only version of "was that a mistake?" that has an answer you can check at the board rather than argue about afterwards.

Walls are graded the same way but with different arithmetic — a wall's error is the gap between the detour it buys and what it costs you in material and turns, which is the whole subject of Wall Economy. What both cases share is that you cannot grade any move without a number, and the number comes from counting both shortest paths every turn, the way Tempo and the Race sets out.

Play Scout with that discipline explicitly: before each move, count both distances; after each move, ask what the best alternative was and by how much you missed it. Against an opponent this gentle the game is nearly a controlled experiment in your own error rate, because very little of what happens to you is being done to you.

How to beat Scout

• Race and count. Do not hunt for tactics. There is nothing to trap here. Scout has no multi-move plan to disrupt, so time spent looking for a clever combination is time not spent being accurate.
• Take the tempo it gives you; do not wait for more. The ceiling of three squares is per move, not per game, and the deliberate-error branch comes up on roughly one move in six, so the give-back genuinely accumulates over forty moves. But more than a quarter of its moves are a plain forward march, so its pawn also makes steady, undistracted progress the whole time. The two roughly trade off. Bank each concession as it arrives — do not deliberately fall behind on the theory that the next blunder will bail you out.
• Answer its walls with your feet, not with walls. Scout's walls come from a shallow search, so they arrive as single walls near a pawn rather than as coordinated pairs. A lone wall usually costs a step or two. Route around it and keep walking; matching it wall-for-wall spends your material to punish something that barely hurt you.
• Do not dump your last walls to "use them up". The moment both of you are out of walls the deliberate errors stop, because the wrapper never sees a move it did not have to search for. Whatever tempo you are down at that point is tempo you keep — Endgames is how you work out, several turns early, whether the race you are steering into is one you have already won. Keep enough material to arrive there ahead.

Do not read much into the final margin. Because the give-backs are random and independent, a game in which several land early looks like a rout and one in which they land late looks close, off the same play from you. The measure that means something is the other one: when you are winning comfortably and can say afterwards which of your own moves were errors and by how much, move up to Mason. You can play the bots as a guest without signing up at all; what a paid plan adds is set out on the pricing page.