How we test
Updated
On this page
The figures on each game page come from boards and test runs we made ourselves; none come from visitors’ play. Each example gives a count and the date of its run. Where our records carry no date, we write “in our test suite”.
Checking that puzzles are fair and solvable
Each game below serves a puzzle only after a program has checked it. Two things are not checked this way: Classic Minesweeper boards, which are random, and FreeCell deal numbers you type in yourself.
| Game | The check | Result |
|---|---|---|
| Sudoku | One solution each, finished with named techniques and no guessing | 2,500 of 2,500 puzzles, 25 October 2026 |
| Nonogram | A separate solver re-solves each picture by row and column logic alone, with no guessing | 249 of 249 pictures, 29 October 2026 |
| Minesweeper | The game’s own solver builds each No-guess board (logic alone can clear it). A separate verifier then re-checked large samples | 30,000 No-guess boards (10,000 per size) and 3,650 Daily boards, all cleared by the verifier, 25 October 2026 |
| FreeCell | Each banked deal has a saved winning line, replayed under the game’s rules | 3,300 of 3,300 lines win, 25 October 2026 |
| Spider Solitaire | Every deal in the bank is one our solver won, and each saved line is replayed on a fresh board | 356 of 356 lines clear eight runs, 25 October 2026 |
| Mahjong Solitaire | Each deal is built by playing the board out. In tests, a separate replay must then empty it | 0 bad deals in 400 (100 per layout), 25 October 2026 |
| Word Search | Every row, column and diagonal is scanned both ways for blocked words (words we keep off the boards) | 3,653 Daily boards and 10,000 per level, 0 blocked words, 25 October 2026 |
Two numbers show why we check. Only 24,976,511 of the 33,554,432 possible 5×5 nonogram pictures (74.44%) can be finished by line logic alone, counted in full on 29 October 2026. In our test suite, plain random filler would put a blocked word on 30.2% of Easy, 48.8% of Medium and 61.8% of Hard Word Search boards (2,000 boards each).
On 25 October 2026, 397 of 400 Mahjong Solitaire shuffle positions were re-dealt and cleared. The game refused the other 3, on the Garden layout, because no clearable shuffle existed.
Replaying Daily puzzles ahead of time
A Daily comes from the game and the calendar date alone, read from your device clock. That means we can build a future date’s puzzle ahead of time, and for the games below we did. What every Daily has in common is on the Daily puzzles page.
- Minesweeper: on 25 October 2026 we built the board for each of 3,650 days, from 1 November 2026 to 28 October 2036. The verifier cleared 3,650 of 3,650, and none repeats within 365 days.
- Mahjong Solitaire: on 29 October 2026 we rebuilt every Daily deal from 1 November to 30 December 2026, 60 in all. Each stored solution still emptied its board, and no deal differed from the build before our edits.
- Nonogram: the same 29 October 2026 run covered 431 Daily dates, from 1 October 2026 (before the first Daily) to 5 December 2027 (Daily #400). A separate solver finished every one, one row or column at a time.
- Fingerprints: a fingerprint is a short code computed from a puzzle, so any change to the puzzle changes it. For six games (Sudoku, Minesweeper, Spider Solitaire, Nonogram, Mahjong Solitaire and Typing Test) we ran one comparison between 25 and 29 October 2026: the fingerprints of the next 60 Daily puzzles, before and after our changes. All 60 matched for each game.
Comparing the rules with a second checker
A game can pass its own tests and still be wrong. So a second method, worked out separately, must reach the same answer.
- Minesweeper, 25 October 2026: 100,000 Classic boards had no mine on the first tile or its 8 neighbors. Our 3BV count (the fewest clicks that open every safe tile) matched a second method on 10,000 random boards.
- FreeCell, in our test suite: the most cards you may move at once matched the formula (empty free cells + 1) × 2 to the power of the empty columns, in all 20 combinations.
Checking scores and timing
Typing Test formula. WPM is the characters in correctly typed words, plus the Space after each, divided by 5 and by the minutes. A wrong word scores 0. On 29 October 2026 a second scorer, written separately, rebuilt 1,000 generated runs from their key-by-key records. Its result agreed with the game’s on all 1,000.
Typing Test timing gate. A run is marked Unranked if it has 30 or more gaps between keystrokes and they are too even: either their spread (standard deviation) is under 3 ms, or at least 8 in 10 gaps fall within 2 ms of the median gap. On 29 October 2026 we tested generated records:
- 10,000 human-like records at 20 to 160 WPM: 0 flagged.
- 3,000 machine records at a fixed 120 ms, with random variation (jitter) of 0, ±0.5 or ±1 ms (1,000 each): all flagged.
- 2,000 records from browsers that round key times to steps of 16.67 or 100 ms: 0 flagged, and all 2,000 noticed as rounded.
These records are made by a program, not typed by people, and the largest jitter we tried was ±1 ms. A bot that copies the uneven rhythm of human typing would get past this gate. We make no claim otherwise.
Checking speed, screens and input
One shared check script runs the same sections on every game.
- Freezes: a freeze, or “long task”, is when the page stops responding for more than 50 ms in one go. On a 390 px screen, with the processor slowed to a quarter of its speed, the script makes 60 random inputs, plays the game’s largest animation and finishes the round. The allowed number of freezes is 0.
- Busy computers: if our computer is over 50% busy before, during or after a run, the result is “not judged”, not a pass. A failure counts only when it was under 40% busy. In our latest runs (October 2026) the computer was too busy each time. So the freeze check ended with no verdict for the seven games in the table above and for Typing Test. We make no speed claim yet.
- Phone widths: at 320, 360 and 390 px, at the start, with the board at its fullest and on the result, nothing may be wider than the screen. Controls must measure at least 44 by 44 px, and dense grids and digit keys at least 24 px.
- No page scroll while playing: on a 390 px screen, a swipe that starts on a live board must not move the page. A swipe that starts in a strip outside the board must scroll it, and that strip must be at least 48 px tall.
- Keyboard: Tab must leave focus inside the game, with a visible ring. Each game’s own script also presses its keys.
- Reduced motion: with the setting on, the script must find no sliding, zooming or spinning between the start of the largest animation and its peak. It cannot catch every kind of movement, so this is a check, not a promise.
- Light and dark themes: the site opens dark whatever your system says, and a footer switch selects light. The script flips themes mid-round, and the round must survive.
These runs use a phone-sized browser window with touch input on. Trials on physical phones and play tests with outside players are not done yet. Typing Test phone input is tested in code with scripted input sequences modeled on iPhone and Android keyboards, not on real phones.
Checking games with computer players
Number Merge, Disc Flip, Bubble Shooter and Reaction Time Test have no single answer to check. Instead, programs play them by the game’s own rules, and we report how far they get. These are programs, not people.
- Number Merge: in our test suite, five bots that never see the next tile played the 365 Daily puzzles from 1 November 2026, and four played 1,000 Classic games each. Three of them were played again on 29 October 2026 and scored the same.
- Disc Flip: each level plays three simple bots or the level next to it. Across 12 matchups, 5,400 games in all, the results of 23 October 2026 were identical when we repeated them on 25 October 2026.
- Bubble Shooter: three bots played 1,000 games each on every level, 9,000 in all, on 27 October 2026.
- Reaction Time Test: in October 2026 we generated 1,000,000 waits and ran bots that press at fixed times, 10,000 sets for each of 20 timings.
What we do not do
- No visitor data in our numbers. Every figure here comes from our own boards and tests, or from a study we name, and the checks run on a copy of the site on our own computer. Our analytics count visits and game events, as the privacy policy explains, but none of those counts are used here.
- No ratings, reviews, leaderboards or accounts. Records and saved games stay in your browser.
- No figure we cannot measure. We give no rate of forced guesses for Classic Minesweeper boards. Our Nonogram picture-naming test used an AI model, not people: it saw 305 pictures drawn as plain grids, and we dropped the 56 it did not name correctly in at least 2 of 3 tries.
Where each game shows its numbers
Most game pages have a section of counts and dates titled “Our numbers”. If a puzzle or a rule looks wrong, tell us. The About page says who runs the site.