What Changed on TheQuoridor.com

This is a note about the state of the site, written to be read by someone who played here a few months ago and wants to know what is different. It is a changelog, not a pitch, and it links rather than repeats — every claim below has a page that owns it properly.

The site now has a written library

Until recently the only way to learn anything here was to play. That has changed. There is now the complete rules — the formal statement of the board, the turn, the jump and the one restriction that keeps a wall from sealing anyone in — and a beginner's guide that assumes you have read the rules and want to know what a good decision looks like instead.

On top of those sits a six-article strategy series: the wall exchange rate that tells you whether a wall is worth spending at all, path counting and the race, the opening, the recurring wall-or-advance decision, endgames where the position stops being a negotiation and becomes arithmetic, and traps and cages — the structures that beat strong opposition.

Two reference pages round it out: a glossary that defines the recurring vocabulary in a line or two each and hands the argument off to whichever page owns it, and an FAQ for the questions people actually send in — do I need an account, which bot should I start with, why did the bot just place a wall that seems to help me.

All of this is plain, linkable pages, reachable without signing in, with a footer on every one of them so you can get from any page to any other. That sounds like a small thing. It is the difference between a site you can read and a site you can only log into.

Seven pages about the opponents

The bot roster used to be a list of names, avatars and numbers. Each bot now has its own page, and they are written to be useful rather than decorative: Scout, Mason, Warden, Oracle, Magnus, Nemesis and the experimental neural engine. The live roster, with current ratings and records, stays where it was: the bots page.

Writing these forced an honest decision. The roster is six difficulty tiers plus one separate experimental engine, and the six tiers are the same program at six different thinking budgets (how the engine scores a move). Seven pages that each said "like the last one, but stronger" would have been seven pages of nothing. So each page carries one lesson that belongs only to it: what an error even is when you are playing a racer; how far ahead a search of a given size can actually see; which phases of a Quoridor game reward calculation and which reward habit; how to find the wall placements worth considering at all; why the move that survives many continuations beats the move that looks brilliant in one; at the top, the structural blind spot that a bigger budget does not fix; and, off the ladder entirely, the difference between a program that calculates and one that judges.

Each page also has a section on how to beat that specific opponent, and the advice differs by tier because the thing you are exploiting differs by tier. Against the gentlest rung of the ladder you are exploiting a deliberately applied handicap (the details). Against the strongest you are exploiting a weakness in what the search values, which is a completely different game.

The engine story, told straight

The playing engine was replaced outright in July. The previous persistent-plan engine was deleted — not disabled, not kept as a fallback — and replaced with a port of an open-source UCT Monte Carlo tree search program, credited in the project's NOTICE file. A frozen copy of the old engine survives in the repository for exactly one purpose: to be the sparring partner the strength harness measures the new engine against. The reason for the replacement is worth stating because it is the kind of bug that hides for a long time: the old engine's position evaluator was, in practice, never consulted, because its simulated games always ran to a finish first. The thing that was actually judging positions was the cheap policy that played those simulations out. Rather than patch around that, the whole search was replaced with a design where the simulations are the judgement, on purpose.

Then came the part worth admitting in public. After the swap, every rung of the difficulty ladder was measured against the rung below it, at each tier's real playing budget, on more than one random seed. The result was worse than "mostly fine". Two rungs separated cleanly. Two came out level — an even split, which does not mean the tiers are identical, only that a match of that length could not tell them apart. And one came out backwards: the top tier failed to beat the tier below it, on both seeds. That is a broken ladder, and the tempting response — call the whole thing seed noise and ship — was available and wrong.

The cause turned out to be dull and specific, which is the good outcome. The per-tier thinking budgets were being resolved through an older pacing table that had been built for the previous engine, and its ceilings were wider than the new ladder's own. At complicated positions the second-strongest tier paced itself up to nearly the same number of simulated games as the strongest one, so the two tiers were separated by a sliver instead of the intended gap. A per-tier ceiling was added on top, the ladder was re-run at both seeds, and the inverted rung came out the right way round. The re-run also reshuffled the middle: one of the level rungs separated cleanly, and a rung that had separated before came out level instead, so the ladder still has two rungs a match of that length cannot split. Both of the ones that moved are reported where they belong, on Oracle's page and Magnus's page, rather than tabulated here.

A regression suite built out of losses

Two games in which a human beat the two strongest bots were recorded move by move and turned into a fixed test suite: every position where the bot was to move and was already deep inside a wall pocket became a test case, pinned with the move it actually played. It is not a suite the engine passes cleanly, and that is the point of keeping it — Nemesis's page explains what it measures and what it says about how to play the top of the roster.

The neural bot is labelled a work in progress because it is one

There is a separate experimental neural engine on the roster. It is not the hardest opponent here. Measured at equal thinking time it currently loses to every tier it has been tested against, including the gentlest rung of the ladder — which makes it the weakest opponent on the roster rather than the crown of it, and the bots page rates it accordingly. It ships anyway, honestly labelled, because a bot that improves in public is more interesting than a bot that appears fully formed. Its page covers what it is, why a program that forms an opinion about a position is a different kind of thing from the six on the ladder, and how it gets better.

More of the site works without an account

Playing a bot and playing a friend on one screen no longer require signing in; both were always accepted by the server and were only ever blocked by the front end, which has been fixed. The leaderboard, the daily puzzles, the position designer and the tournament calendar are now open to signed-out visitors too. The leaderboard and the calendar are things you read; the designer is not — a signed-out visitor can build a position on it and play the position out, and only the actions that genuinely need an account raise a sign-up prompt, in place of the silent redirect that used to happen.

Two details that catch people out. Bot games run on a clock, but a pass-and-play game on one screen does not: single-client modes are deliberately pressure-free, so there is no deduction and nothing to flag on. And neither is rated — playing bots cannot move your number in either direction, whoever you are. If you make an account later the games you already played can come with you. The FAQ covers how ratings, time controls and guest games interact, and Pricing covers what a paid plan adds on top of the free game.

Cosmetics

The board picked up a proper cosmetics layer over the last few weeks: pawn skins, per-player and per-owner wall colours, selectable board styles, and a glow option — with the pieces you have equipped now following you into every game mode instead of only the one you bought them in. Saved replays and exported clips render the cosmetics the game was recorded with, so a replay looks like the game did rather than like your current loadout.

What is still open

Three of the things above are unfinished rather than shipped, and collecting them in one place is more honest than leaving them scattered through a changelog. The suite built out of the two lost games is still not passed cleanly. Two rungs of the ladder still come out an even split rather than a clean separation, and whether that is two tiers genuinely sitting close together or a match too short to tell them apart is not something the run answers. The experimental neural engine is on the roster while losing to every tier on it.

They are written down so that the next note touching any of them has to say which way it moved — the blog index sets out what a note has to clear before it goes up.