Most puzzle sites ask you to take their word for it. This page is the alternative: what actually happens to a grid between being generated and being handed to you, which of those steps are automatic, and where the honest gaps are. If you ever find a puzzle that breaks one of the promises below, that is a bug worth reporting rather than a puzzle worth persevering with.
Nothing is served until it is proved
A puzzle starts as a complete, valid grid. Digits are then removed one at a time, and after every single removal the remaining grid is handed to a solver that counts how many ways it can be completed. If the answer is anything other than one, the digit goes straight back. That check is not a final pass at the end — it runs on every candidate removal, which is what makes the guarantee hold rather than usually hold.
The same applies to everything drawn on top of a grid. A cage layout, a set of dots, a fleet, a map of rings and frames: each is verified under the exact clues you will be shown, not under the ones the generator happened to start from. A puzzle that arrives by a shared link is re-proved in your own browser before the page will play it, because a link is just text and text can be edited.
Why some of these searches are deliberately capped
Proving that a grid has no answer can take longer than finding one — on an irregular layout the solver may have to exhaust the whole tree to be sure. So the searches that could run long are given a step budget, and running out of budget counts as failure rather than success: the clue stays, or the puzzle is discarded. It is the conservative direction, and it is the reason a shared link can never be made to hang the page.
Difficulty is measured, not asserted
A label like "Expert" usually means a clue count, and clue count is a poor guide — a grid with twenty-two digits can be gentler than one with thirty. Here the label is decided by solving the puzzle the way a person would. A registry of human techniques is applied in order of difficulty, cheapest first, and the puzzle is named after the hardest technique the solve actually needed. If a grid can be finished by looking for squares with one candidate left, it is Easy, however few digits it started with.
That machinery is also what the Hint button uses, so a hint can name the move rather than hand you a digit, and what the two hardest banks were built with. Ten thousand puzzles were each replayed by the grader and sorted by what they genuinely demanded:
Those last few are the point of doing it. Nearly one puzzle in five arrived carrying a harder name than it deserved, and re-grading moved it. Re-running that curation has to reproduce the published banks byte for byte, and that is treated as a test rather than a convenience: if a change to the technique registry quietly alters what "Evil" means, the files stop matching and we find out immediately.
A variant's extra rule has to earn its place
Adding a rule to sudoku is easy. Making it matter is not. It is entirely possible to generate a board with diagonals, or cages, or shaded squares, where the extra rule never actually constrains anything — the ordinary sudoku clues settle the grid on their own and the decoration comes along for the ride. A player who ignores the rule and finishes anyway has caught the setter out.
Strip a variant's extra rule out and its puzzles must stop having a single answer. If they don't, the rule was decoration and the puzzle is discarded before anyone sees it.
That test is applied per puzzle, at every level, and it rejects real work rather than rubber-stamping it: dug at forty clues, more than half of the attempts on the windowed board fail it, and better than a third on the knight's-move board. From the harder levels upward there is a second requirement — the graded solve has to actually fire one of the variant's own deductions, not merely tolerate the rule being present. Some boards need a ceiling as well as a floor, where marking too much can quietly take the puzzle apart; the pages that apply to say so themselves.
What gets checked that you can't see
The puzzles are only half of it. Every route on the site is run through an automated accessibility audit in both the light and dark themes, and the checks that a machine cannot settle — colour over a gradient, mostly — are measured by hand instead. No colour is added to this site until its contrast has been calculated against every surface it can end up on, including the tinted states a square takes when it is selected, related, hinted or wrong. That arithmetic has changed shipped designs more than once: the parity marks on the odd-even board are hollow outlines rather than the filled pads most published puzzles use, because the filled version put entry ink under the accessibility threshold on a dark board.
Alongside that, and before anything is published:
- The whole game is playable from the keyboard — digits, arrows, notes, undo and redo — and every control is reachable in a sensible order with a visible focus ring.
- Touch targets on phones are at least 44 pixels on both sides, which is stricter than the accessibility guidelines require and closer to what a thumb actually needs.
- Pages are checked for horizontal overflow at phone, tablet and desktop widths, and the board is sized so that the grid and the number pad are both reachable without hunting.
- Motion respects the reduced-motion setting, so nothing slides or lifts if you have asked your device to stop that.
- An automated suite — 318 checks at the time of writing — has to pass before any change ships. Most of it is about puzzles being solvable and correctly graded; the rest guards the writing, including a test that fails if any two pages ever share a sentence.
What we don't claim
Three honest limits. The generator proves that a puzzle has one answer and that logic can reach it, but "solvable by logic" is measured against our own registry of techniques — a puzzle we call Extreme is one our ladder cannot finish, which is a statement about the ladder as much as the grid. Performance is watched but is not yet part of the automated suite: layout stability and load times are measured by hand in a browser rather than audited on every change.
And there are no accounts here, so times, streaks and saved games live in your browser and nowhere else. That is deliberate — it is why the site can promise it does not hold your solving history — but it does mean clearing your site data clears the record, and nothing follows you to another device. If any of this changes, it changes on this page first.
Questions about a specific puzzle are the most useful kind we get. There is a contact page, and if you would rather read about the people doing the work than the process, the about page covers that.