Intermediate — Quoridor AI Opponent

Mason is the intermediate bot, the second rung of the difficulty ladder (the bots page has the current roster and ratings). It is the right opponent once you have stopped losing games to Scout and can count a race without stopping to think about it, and it is the tier where a specific skill starts paying: building a plan whose value only arrives on the second wall, and playing the moves in the order that hides it. That skill is the whole subject of this page, because Mason's search is small enough that its horizon is a real, exploitable feature of the board — not an abstraction.

Its horizon, described honestly

Mason's allowance is a single flat cap: a playout count in the hundreds rather than the thousands, plus a short time allowance, and whichever runs out first stops the search (engine/mcts/engine-v3.ts:298-308, with the cap coming from the low-tier lerp at engine/elo.ts:82; see how the engine scores a move). Flat is the operative word. The per-position envelope, which lets a tier think longer when the board is complicated, does not exist below advanced (engine/bot-tuning.ts:51-53 returns nothing under that tier; how complexity sets a budget is where the envelope is explained). So a tangled midgame with four walls down and both pawns committed gets exactly the same thinking as a step down an empty file. The moment the position becomes hard is the moment it stops getting extra help.

A search that size is not "shallow" in a vague sense. It is shallow in a countable one, and the count is worth doing once, because everything further down this page follows from it.

Start with how the tree grows. A node does not get children the first time the search reaches it — it gets a single playout, and only on its second visit does it expand and generate its moves (engine/mcts/uct-search.ts:65-74). And the root has a great deal on offer. It generates the pawn's steps — all of the legal ones while you still hold walls, and only the forward ones once you do not — plus, whenever Mason still has walls of its own, a stencil of candidate wall placements (engine/mcts/uct-search.ts:106-137). In a normal midgame that stencil is the bulk of the list, and the list runs to dozens of entries.

Now divide. A few hundred playouts spread across dozens of root candidates leaves each candidate sampled a handful of times. The division is not even — the search keeps returning to whichever candidates are scoring well, so the best few do get visited enough to expand and be explored past your reply, and the rest are visited once or twice and abandoned. But the sample size is the ceiling on what the tier can resolve, and here that sample is small enough to name: a difference that only shows up after two more accurate moves by both sides has to survive a handful of random finishes to be noticed at all.

Hence the practical translation. Mason resolves the consequence of this move — a wall that costs it more than it costs you, a step that walks into a jump, a race it can win by moving now — because those differences appear in the playout immediately and in nearly every sample. It does not reliably resolve the consequence of a move you have not made yet. Anything whose payoff arrives two placements later has to be found by chance rather than by search, and at this budget chance does not get many attempts. Mason answers the board in front of it: well against a threat that already exists, poorly against one that is still being assembled.

One difference from the tiers above shows up at the board and follows from the same gate. Mason does not switch into a straight race when its own walls run out — that shortcut starts at advanced — so it keeps searching, and keeps being weakened, in positions where walking forward was its only real option.

How to beat it: sequence the plan, do not announce it

Against a stronger opponent, the case against a slow wall structure is a tempo case. The early walls of such a structure are worth nothing on their own, so the defender banks a step for every wall the builder spends, and can often answer the whole plan by simply walking — traps and cages does that arithmetic in full, and it is right against anyone who is looking.

Against Mason the arithmetic does not change. What changes is the looking. The single step that defuses the plan has to be found, and finding it means resolving what your next wall will do to a position that does not exist yet. That is precisely the thing this budget does not reach. The tempo you spend on a quiet first wall is still really spent — count it — but you are far more likely to be allowed to place the second one. That is the specific licence this tier gives you, and it is worth taking deliberately rather than by accident:

Pick the pair before you place either. Decide on both walls first, and check that the second one is still legal and available after the first goes down. A pair chosen in advance beats a wall placed hopefully — the failure mode of this plan is committing to the first and then discovering the second slot was never there.

Lead with the wall that costs nothing. The anchor — the one that adds zero to Mason's path today — goes first, leaving the painful wall for last. A wall that changes neither distance gives the search nothing to register: the position it evaluates after that wall scores the same as the position before it. Note what is not being claimed here. That wall is not harmless, and a human opponent should read it as the loudest warning on the board; traps and cages lists it first among the tells, because a wall worth zero next to your pawn is what a fence post looks like from the inside. The claim is only that Mason will not read it that way.

Wait for commitment. These plans work against a pawn that has already gone deep and to one side, where the board edge does half your work. If its pawn is still central and flexible, the pair you were planning is worth much less; walk instead and re-check in three moves.

Count first. Two walls and two turns is a real investment, and a plan that lands but does not change who arrives first is a loss with extra steps. Do the count before you start, with the detour included — tempo and the race is the page for that arithmetic.

Do not fear complications. A denser position genuinely is harder for Mason, because its allowance will not grow to meet it. That is not a licence to play badly on purpose; it means that when a sharp line and a quiet line are close in value, the sharp one is worth more against this tier than the count alone suggests.

The symmetry matters too. Mason is as unlikely to build a two-wall structure against you as it is to defuse yours, for the same reason: the payoff sits past its horizon in both directions. You can walk a slightly more committal route here than you would against the tiers above, provided you still check the square in front of you the way any decent player does.

The moves it gives away

Like Scout, Mason still runs the site's deliberate weakening wrapper outside its search — the same mechanism, dialled down. One consequence of the dialling is specific to this rung. The wrapper's deliberate-error branch only accepts a move that concedes at most its tier's magnitude, and by intermediate that ceiling has fallen below two squares (engine/mcts/weakening.ts:91-92, filtered at :125-129, with the tier anchors at engine/elo.ts:4-11). Since path lengths come in whole squares, the only mistake that still qualifies here is a one-square give-back. Scout can hand you three; Mason can hand you one.

Read those moves correctly. A weak-looking move from Mason is usually not a plan you have failed to spot — it is the tier. Take the square and carry on with your own plan rather than searching for the trap that is not there. The glossary has the vocabulary if any term here was new.

When to move up

Move up when you win the majority of your games against Mason and — this is the better test — when you win them because a plan you chose in advance worked, not because it handed you a square. The next rung, advanced, is the first tier with no weakening applied to its choice at all. The bots page has every opponent and their current rating, and the site's plans cover what comes with each one.