Seventeen is a floor, not a record
A clue is a digit printed on the grid before you begin, and a puzzle is proper when its clues leave exactly one way to finish. Ask how few clues a proper puzzle can carry and the answer has been settled since 2012: seventeen. Not seventeen as the lowest anyone has managed, but seventeen as a wall. Sixteen has been shown to be impossible, and the puzzles that sit exactly on the line have all been found and counted.
The impossibility was established by Gary McGuire, Bastian Tugemann and Gilles Civario, who did not reason their way to it so much as look. There are 5,472,730,538 essentially different completed grids, and their program examined every one for a set of sixteen squares that would pin it down alone. About 7.1 million processor-hours later, none of them had one. It remains a computation rather than an argument anyone can follow on paper, and that is worth admitting; but a computation that visits every case is a proof all the same.
How you search five billion grids for something that is not there
Every completed grid contains small groups of squares whose digits could be swapped around into a different completed grid — two 1s and two 2s at the corners of a rectangle spanning two boxes are the simplest case. A clue set that misses such a group entirely cannot tell the two grids apart, so a sixteen-clue puzzle would have to touch every one of them. The search listed those groups for each grid and asked which sets of sixteen squares hit them all; the few that did were then checked outright, and every one of those admitted a second answer.
Where the 49,158 come from
Gordon Royle, a mathematician at the University of Western Australia, kept the list for years, growing it from puzzles sent in and puzzles found by search until it stood at 49,151. Others took it on and pushed it to 49,157, and the last, number 49,158, turned up in 2019. Then the question was closed from the other side: a scan of every completed grid, run the way the sixteen-clue search was run, found nothing the list did not already hold. So what this page serves is not a sample of seventeens. It is all of them.
Each entry stands for a family rather than a single grid. Rename the digits, swap the three rows of a band among themselves, swap whole bands, do the same with columns and stacks, reflect the lot across the diagonal: every one of those moves turns a proper puzzle into a proper puzzle with the same solving path, and together they give one entry up to 1,218,998,108,160 forms. The list keeps a single form per family and chooses it by sorting — the version that reads earliest as a string of characters is the one written down. Reading earliest means opening with as many empty squares as possible, which is why the whole file begins with a run of dots and every stored puzzle has its clues crowded into the lower right.
The board never shows a stored form. Each press of New puzzle draws a rearrangement at random, so the grid in front of you has almost certainly never been printed anywhere. That could only be honest if difficulty survived the shuffle, so it was checked: two thousand puzzles were graded before and after a random rearrangement, and the band and the hardest step came out identical in every one.
A rearranged sudoku is the same sudoku. The list keeps one copy per family, and this board shows you any of the trillion it stands for.
Seventeen clues is not seventeen times harder
The reputation is fearsome. Search for these puzzles and you will read that they take hours, that they need forcing chains, that most people cannot finish one without a computer. Nobody writing that had graded them, so we did: every one of the 49,158 was replayed with the same ladder of human techniques that grades every board on this site, and sorted by the hardest move it needed.
45 per cent of them fall to hidden singles alone. Add the puzzles that need nothing beyond a subset or a locked candidate and 85 per cent of the collection sits at Hard or below — the level where pencil marks stop being optional, and nothing more exotic is asked for. The wings, fish and colouring of Expert account for 11 per cent; the almost locked sets and swordfish of Evil for under two; and 1,299 puzzles, 2.6 per cent, run past the end of the ladder into chains or a considered guess.
The hardest step tells the same story from the other side. Hidden singles top the table because a puzzle that needs nothing else is graded by them; then come pointing pairs, then the skyscraper, and everything below that is scarce.
| Hardest step needed | Puzzles | Share |
|---|---|---|
| Hidden Single | 21,905 | 44.6% |
| Pointing Pair | 14,353 | 29.2% |
| Skyscraper | 2,817 | 5.7% |
| Naked Pair | 2,383 | 4.8% |
| Hidden Pair | 1,833 | 3.7% |
| Beyond the ladder | 1,299 | 2.6% |
| XY-Wing | 1,180 | 2.4% |
| Box/Line Reduction | 1,117 | 2.3% |
| Everything else | 2,271 | 4.6% |
"Everything else" gathers the nine rarer endings: W-wings, almost locked sets, colouring, triples, the fish, and the two puzzles whose hardest move is a jellyfish.
So where does the reputation come from? Partly from the 1,299 that earn it. Mostly from what sparseness does to the experience rather than the logic. With sixty-four squares to fill and only three or four hidden singles on offer at the start, the opening is slow and unforgiving of a missed pencil mark, and on a Medium seventeen every one of the sixty-four steps is a placement, with no easy stretch to coast through. That feels harder than a Hard puzzle with twenty-six clues. Measured by the moves it asks for, it usually is not.
The number of clues measures how much is missing, not how hard it is to find.
Why none of them is easy
On this site Easy means one thing: the whole grid falls to naked singles, squares with a single digit left. Not one seventeen qualifies, and the reason is arithmetic. A square has twenty peers, and it becomes a naked single only once eight different digits are standing among them. Seventeen clues rarely aim that many at one place: 88 per cent of the collection offers no naked single at all before the first move, 5,719 puzzles offer exactly one, and 49 offer two.
The opening move, then, is a hidden single: a digit with one home left in some house. Every seventeen but 49 has one available before anything is placed, and on average only 3.4 are on the board at once, which is why the first minutes feel like hunting for a needle. Just 35 puzzles offer no single of either kind at the start; on those the first move has to be an elimination, a pointing pair that thins the candidates before any digit can go down.
What seventeen clues have to look like
A set of clues this small cannot take any shape it likes, and one constraint can be reasoned out before looking. A proper puzzle can leave out at most one digit. Were two missing from the clues, take any solution and swap those two digits throughout it: every row, column and box still holds each digit once, no clue has been contradicted, and you hold a second solution. So a seventeen may omit one digit entirely, and nearly half of them do — 23,788 of the 49,158 use only eight. They run a shade harder than the rest, 13 per cent of them grading Expert against 9 per cent of the nine-digit puzzles, though not by anything like the margin the bare grid suggests.
The rest of the shape was measured rather than reasoned:
| Across the collection | Puzzles | Share |
|---|---|---|
| A row or column with no clue in it | 32,228 | 66% |
| A whole box with no clue in it | 4,596 | 9% |
| Two empty boxes | 71 | 0.1% |
| Only eight digits used | 23,788 | 48% |
| Fullest box holds 3 clues | 42,590 | 87% |
| Fullest box holds 5, the most any has | 6 | <0.1% |
An empty line is not a weakness. The puzzle is proper, so those nine squares are fully decided by clues that sit elsewhere; two thirds of the collection has one, and any of them can be turned so that the bare line runs down a column instead.
Five levels from one number
Because the whole collection is graded, the level buttons above the board mean here what they mean on the classic board: Medium serves a seventeen that falls to hidden singles, Hard one that needs a subset or a locked candidate, and so on up. Each band lives in its own set of files and a press of New puzzle fetches one slice of one band, so the page never downloads the four-megabyte collection to hand you one grid. The line above the grid names the actual hardest step of the puzzle you were given, not the band's general promise.
There is no Easy button, and not because it was left off: there is nothing to put behind it. Extreme, by contrast, has 1,299 puzzles behind it, grids the ladder cannot finish, where the extreme page's case about chains and testing applies in full. Evil has 814, decided almost entirely by almost locked sets, and the 5,398 at Expert are the place to meet a skyscraper in the wild, since it is the commonest expert-level ending in the collection by a distance.
Proved twice
Uniqueness was established two ways. The whole collection was counted outright once, every puzzle searched for a second answer and none found. The curation that builds the banks then proves it again on every run, by a shortcut worth spelling out: every technique on the ladder except the unique rectangle is direct logic, removing only what cannot be part of any solution, so a puzzle the ladder finishes without that one move has been shown to have exactly one answer simply by being solved. 47,425 of the 49,158 are proved that way; the 1,301 the ladder cannot finish, or finishes with a unique-rectangle step, are counted the slow way, and so is a one-in-a-hundred sample of the rest, to keep the shortcut honest.
Playing a seventeen here, and on paper
Sixty-four blanks make pencil marks a necessity rather than a habit, and the Fill button writes the complete set in one step so the opening scan can begin from a full picture. From there the natural rhythm is digit-first: choose a digit, ask each box in turn where it can still go, and place it wherever the answer is one square. The medium page teaches that scan in full, and a seventeen is the purest exercise in it there is, because for most of the collection it is the entire solve.
The share button copies the 81 squares into a link, so a rearranged form that has never existed anywhere can still be sent to someone and opened exactly as you saw it; a game reached that way is saved under its own key and never overwrites the one in progress here. The pack builder above puts the collection on paper, each grid freshly rearranged, with its band printed beneath it and the answers at the back. It stops at four to a sheet, because a grid with sixty-four squares to mark up wants the room.