By Nick
Slidesmith is our third game and the first one we invented. Star Battle and Queens are old puzzles that we generate well. Slidesmith started as a toy on a shelf, spent a while as the wrong game, and had one serious flaw that only showed up once other people played it. This is how it got from there to a walled board with a par on it.
It started with a board game
The idea came from a two-player board game. Each player has a five by five tray of coloured tiles with one empty space. You shake a box of nine small cubes to get a three by three pattern, then race to slide your tiles until the centre nine match it. The slide is the good part: push a tile and the whole line between it and the space moves with it.
Underneath, it is the oldest puzzle on any phone: the fifteen puzzle, the one with the numbered tiles and the missing corner. Those are fun for about as long as it takes to reach the last row. Then every slide puzzle ends the same way, cycling three tiles round a corner until the last one drops in. A fixed target zone brings that endgame with it.
Version one: sixty seconds
The first Slidesmith was a speed game on a personal site. Eight small patterns, sixty seconds, a timer counting down. It was a decent way to lose a minute, and it had the rule that survived everything since: the pattern is small, and it counts anywhere on the board. No target zone, no last-tile shuffle. You look for the place where the pattern is nearly made already and you finish it there.
What the clock did was reward tapping. The fastest way through eight patterns is to slide first and look second, which is the opposite of what makes a slide puzzle interesting. Nobody finished a game wondering whether there had been a shorter way.
The turn: slide-puzzle golf
So the clock went, and the question changed from "how fast?" to "how few?". A run is five patterns, built one after another on one board that is never reset. Your score is the number of moves. Before the run goes live, the server plays it too, every round in the fewest moves, and the total is published up front as par. It never changes. Fewest moves against par, on a course you can replay as often as you like: slide-puzzle golf.
One rule does most of the work. A tap moves every tile between it and the gap, and it is still one move.
We prototyped the board game's fixed centre zone first and threw it out, for the endgame reason above. Anywhere-matching with five rounds on a carried-over board was the version worth building. It went live as a beta for signed-in players, three runs a day: easy, medium and hard.
What beta found: you could play perfectly and lose
The complaint arrived in its purest form from the person least able to ignore it. I played an easy run, took every round in the fewest moves the game itself said were possible, and finished three over par. The card marked down rounds I had played perfectly, and a replay would not have helped: nothing on it said which finish had been the wrong one.
The cause was a tie that nobody could see. Here is a round with a one-move finish.
Both finishes cost one move. Both are correct. But the boards they leave behind are different, and the next pattern was this one:
From the second board it is 2 moves away. From the third it is 3. Par is worked out one round at a time, and when a round has several shortest finishes it quietly keeps the one that leaves the next round cheapest. A player has no way to know which that was. Pick the other one and you are a move behind for the rest of the run, having done nothing wrong.
We measured it. We simulated players who never waste a move, taking every round in its true minimum and choosing at random between equally short finishes. On easy runs, about two in three of them missed par. One run could not be parred that way at all. Meanwhile real newcomers were finishing easy runs far over par and some were not finishing them, and "play it again for a better score" is poor advice when the replay cannot show you where the moves went.
The fix: no wrong finishes
We did not change what par means. We changed which runs get published.
When the generator picks the next pattern for a run, it now works out every board a fewest-moves player could be standing on at that point, not just the one par is standing on. The pattern is only accepted if it costs exactly the same from all of them. That is carried through all five rounds, and the finished run is checked exhaustively before it goes in the bank. We also stopped the same pattern turning up twice in a run, which used to read like a bug.
The result is a sentence we can now print: there are no wrong finishes. Play every round in the fewest moves and you land on par. The marks on the result card changed to match. A round is green when you played it in the fewest moves possible from your board, not from par's.
For a while hard kept the tie on purpose. The argument was that choosing between equal finishes is a skill, and the Next preview makes it possible. We dropped that. A cost the player cannot see is not difficulty, on any tier, so every run is now built to the same rule. The tiers differ in the job they give you: spot it on easy, where each pattern is nearly made and a slide or two away; build it on medium; plan it on hard, where the shapes are four tiles and the later rounds take three or four moves to find.
Then we built walls
Golf needs a course. An open five by five grid is a driving range: every day's board had different colours in different places, but no day's board had a shape. We wanted something on the board the eye could hold on to, the way the thick region borders do in Star Battle and Queens.
So each run now has walls: short ink bars on the lines between tiles, fixed for all five rounds. The rule is one sentence. Walls work like the edge of the board: nothing slides through one.
The first walled version was worse
The first walled runs made every tier much harder, which was the opposite of the point. The reason was simple once we saw it. We had drawn the walls, worked out par on the walled board, and published. But a wall across the short finish does not make a round more interesting. It makes it longer. A pattern that was one slide away on the open board could be several slides away round the end of a bar. Par went up to match, and the easy run stopped being easy.
So walls got a rule of their own. A run is only published if every round's shortest finish is as short with the walls as it would be with the walls taken out. The generator solves each round twice, once on the walled board and once on the open one, and throws the pattern away if the numbers differ. We also use fewer walls than we started with: a wall or two on easy, a few on medium, the most on hard.
What is left is what we wanted in the first place. Walls cut away options and never stand in the way of the best one. With fewer tiles able to move there is less to weigh, and the bars give the eye something to steer by: "the column beside the long bar" is easier to find again in round four than "column three". If the only line you can see runs into a wall, that tells you something too. It is not the short one.
A few things walls do not do. They never freeze a tile: bring the gap round and everything behind a wall can move again. They have no say in matching, so a pattern can sit straight across one.
And they cannot strand you. Every slide can be reversed by one slide, and the walls never move during a run, so any board you have been on you can get back to. That is why the layout is fixed per run and not per round. We tested walls that were redrawn each round and they were worse in every way we measured.
We also tested whether walls alone would have fixed the par problem. They would not. Scattering walls on the board barely moved the share of perfect players who missed par. The fairness rule does that job, and it runs on walled boards just the same. Walls are there for a different reason: they make each day's board a place.
How a course is drawn
- The walls are chosen for the run before any pattern is, and never adjusted to suit a round.
- A few connected bars, not a scatter of posts. The eye counts three shapes faster than seven.
- At least one bar touches the rim, so the layout reads as part of the board.
- No closed rooms, and no cell boxed in on three sides. A dead-end pocket turns out to trap its tile for good.
- Bars do not touch each other, and none of them is allowed to make a round longer.
The layout stays the same through every replay, and replaying is the point. Only your best run counts. The second attempt at a course you have already walked is where par starts to look reachable.
What stayed the same
Five patterns, one board. One tap is one move, however many tiles it carries. Par is set before you start and never changes. Only your best run counts. The rules fit on a card, and the full version takes two minutes to read. When the card says forty and par says fourteen, there is a page for that too.
Today's three runs are here. Start on easy, and walk the course first.