Reversi · the computer, measured

Sixty-four squares.
Twelve opponents, each one measured.

Reversi on the classic green board: place a disc so that it brackets a line of your opponent’s, and every disc in the line turns to your colour. The computer thinks in a background worker with an alpha–beta search, and in the last squares of a game it stops estimating and plays the ending out exactly. Every level on the ladder is free to play, the strongest as much as the first.

The levels are not adjectives. Each one is two dials — how far ahead it looks, and how much random noise it adds to its own judgement — and each was scored on the same 64 seeded positions by how many discs a move it gives away to the deepest search on the page. The rungs were then picked by those numbers, and the numbers are printed below, beside every level.

The board

Your move.

Black moves first. The pale dots are your legal moves and the orange one marks the last move played. If you have no move the game passes for you and says so; it ends when neither side can move.

Reversi

Loading the engine…

2 Black You
2 White Computer

Level

You play

The game so far

    Keyboard: the arrow keys move around the board, Enter or Space plays the square; H asks for a hint, U undoes your last move, N starts a new game.

    The ladder

    Twelve levels, chosen by the numbers.

    Every level is the same engine turned by two dials. Depth is how many moves ahead it searches before trusting its evaluation, and the depth also sets where it starts solving the ending exactly. Noise nudges each candidate move’s score by a random amount, up to that many discs either way, before it picks the highest — so a noisy level mostly blunders when moves are close, and a move far worse than its best can never be chosen.

    To score a level, the page’s deepest search — depth 10, exact from 16 empty squares — proved the value of every legal move in 64 seeded positions. A level’s loss is the average number of discs a move it gives away against that yardstick. Lower is stronger.

    The measured ladder: each level's two dials, the discs a move it gives away to the yardstick, and its games against the level below
    LevelDepthExact fromNoiseLoss a moveStepBeat the level belowvs below
    112±647.75——
    212±205.85−1.9069% of 24
    312±104.36−1.4969% of 24
    424±103.38−0.9875% of 24
    5612±102.60−0.7896% of 24
    6816±101.87−0.7265% of 24
    7916±61.16−0.7175% of 24
    8916±50.94−0.2248% of 24
    9916±30.52−0.4281% of 24
    10916±20.34−0.1854% of 24
    111016±1.50.15−0.1975% of 24
    12yardstick101600.00−0.1573% of 24
    • The yardstick is the top level: depth 10, exact from 16 empty squares. It proved the score of every legal move in 64 positions, and a level’s loss is the best of those scores minus the score of the move it plays. The top level scores 0 because it is the yardstick.
    • The positions are seeded random playouts (seed 5440928), eight at 50, 44, 38, 32, 26, 20, 16 and 12 empty squares, with at least four legal moves (9.9 on average). In the 16 with 16 or fewer empties the yardstick solves exactly, so the loss there is in true discs of final margin.
    • For scale: a player choosing among its legal moves at random gives away 9.12 discs a move; Level 1 gives away 7.75.
    • The rule that picked the rungs: walking up, depth never falls, noise never rises, the loss strictly falls, and every step is at least two standard errors on the paired positions. Among the ladders that obey it, the one closest to targets spaced evenly on log(loss + ½ disc). Every step on the ladder above clears that bar; the smallest is 0.15 discs.
    • The noisy levels’ losses are computed exactly, not sampled: the chance of each move winning the noise draw is integrated. As a check, the engine’s own noisy play was sampled 7,200 times at three settings — depth 2, noise ±8: 3.66 computed, 3.70 sampled; depth 5, noise ±3: 1.79 computed, 1.80 sampled; depth 8, noise ±1: 0.23 computed, 0.23 sampled.
    • Then end to end, in real games: every level played the level below it 24 times, from 12 seeded openings with the colours swapped. The upper level won on points in 10 of the 11 pairings and on average margin in 11. Level 8 scored 48% against Level 7 while winning by 5.3 discs a game on average: two levels 0.22 discs a move apart are too close for 24 games to separate, and the paired measurement over 64 positions, which can, puts that step at 6.8 standard errors.
    • Speed, on the machine that measured it (a desktop Mac, JavaScriptCore): the top level’s median move took 114 ms, one in ten took over 361 ms, the slowest 0.6 s. A phone is slower; the moves are the same.
    • Measured 2026-09-28 on the engine whose sha256 begins 9ef3196de5dc42b7. The page’s test refuses a ladder measured on any other engine.
    Every candidate measured — 10 depths by 18 noise settings
    Average discs given away a move, by depth (rows) and noise (columns); the chosen levels are marked
    Depth0±0.5±1±1.5±2±3±4±5±6±8±10±12±16±20±24±32±48±64
    13.183.253.323.363.383.453.523.613.724.014.36L34.705.355.85L26.236.767.397.75L1
    22.232.412.482.532.572.652.732.832.933.143.38L43.634.164.635.085.796.647.13
    32.322.332.352.412.482.572.692.832.983.293.603.874.384.815.195.856.657.12
    41.431.551.671.821.952.142.292.432.582.903.213.514.064.534.915.586.426.93
    51.511.461.481.581.681.872.062.272.492.883.243.554.074.504.865.496.336.84
    61.011.161.181.241.331.531.691.831.982.292.60L52.893.443.944.355.035.976.54
    70.771.071.131.201.281.481.671.862.042.402.733.023.543.994.364.995.916.47
    80.140.150.200.260.330.510.700.881.071.481.87L62.222.833.363.814.595.606.21
    90.190.170.210.260.34L100.52L90.720.94L81.16L71.591.992.332.923.453.894.625.596.19
    100.00L120.020.080.15L110.240.450.640.810.971.331.712.042.633.163.624.415.436.06

    Each cell is the average discs a move that depth and noise give away to the yardstick over the same 64 positions. Read along a row and the loss mostly climbs with the noise; read down a column and it mostly falls with the depth, but not always: an odd depth often measures a shade worse than the even depth below it, against an even-depth yardstick. That is why the rule above lets the numbers, not the depth, choose.

    Method

    What is computed, and what is not.

    The rules

    Standard reversi, as the World Othello Federation plays it: an eight-by-eight board; four discs to start, white on d4 and e5, black on d5 and e4; black moves first. A move must bracket at least one straight line of the opponent’s discs — across, down or diagonally — and every bracketed line turns over. A side with no legal move passes; the game ends when neither side can move. The final count gives any empty squares to the winner, which is the Federation’s rule and the one the engine scores by.

    The engine

    The board is held as two 64-bit sets, and legal moves are found for every square at once. The search is negamax with alpha–beta pruning and principal-variation windows, iterative deepening, move ordering (the previous iteration’s best, then corners first and the moves that leave the opponent fewest replies) and a transposition table that stores whole positions, so a collision can cost time but never a wrong score. The table is emptied before every move, so a level never borrows a deeper search than the one it was measured at. In the last squares an exact solver takes over and plays every line to the final disc. It all runs in a Web Worker, off the page’s own thread.

    The evaluation, and its unit

    Before the end, a position is judged on mobility, frontier discs, corners, the squares that give a corner away, stable edge discs, parity and — late — the disc count, each weighted by how many squares are left. The whole is scaled into discs of final margin by one fitted constant, 1.04: on 180 seeded positions with 16 to 20 empty squares, each solved exactly, that is the slope of the true margin on the raw evaluation, and the two correlate at 0.74. The plain disc count, for comparison, correlates at −0.19: having more discs in the middle of a game is, if anything, bad news.

    What it cannot say

    The yardstick is itself a search, not perfect play: before the last 16 squares its scores are estimates in discs, and exact only after. The top level scores zero because it is the yardstick — a property of the ruler, not a claim of perfection. For the whole game the answer is known: Hiroki Takizawa’s computation of 2023 found that perfect play from the first move ends in a draw, the most a person or a super intelligence could take from a side that never errs. The test positions come from random play, not from the openings people choose. The times are the measuring machine’s (a desktop, JavaScriptCore); a phone is slower, and the moves themselves are the same.

    What leaves your browser

    The game itself sends nothing: no move, no level and no record. The rules, the search and the ladder’s numbers are files served with the page, and the computer thinks on your device. The game in progress, your level, your colour and your record against each level are kept in this browser’s local storage, under labs.reversi and labs.reversi.record, and nowhere else; . The site’s analytics (Google Analytics, on labs.llc only) count the visit, that the board or its controls were used, thirty seconds of engagement and how fast the page loaded — never a move.