Run the Gamut

Under the hood

Every position here is a real solver-verified snapshot from actual self-play, not a hand-set-up puzzle -- each game's own harvest script plays out games against itself, takes candidate snapshots, and only keeps the ones where the exact win/loss/draw solver already has a proven answer. Distances are deliberately kept shallow: a "forced win in 11 more moves" position might be correct, but it's not something you can size up in a few seconds glancing at a frozen board, and a few-second glance is the whole premise of this format. Most games cap at 4 plies to the forced result; Nim is the one exception, left uncapped, since a pile of sticks with a proven parity answer doesn't get harder to read just because the countdown is longer -- there's no board shape to visually parse the way there is everywhere else.

Pacing was tuned from a real timed run, not guessed upfront: the clock allows 8 seconds per round (up from an original 6, after an actual run came back "didn't make it through all seven" at the tighter number), with a flat 5-second penalty added to the clock for every wrong call. The whole run's time budget scales with however many games are actually in a given day's set, with a floor so a short set never turns into an unplayably fast sprint.

The share image reuses the same board-rendering the recap thumbnails already use -- a fresh off-screen copy of each game's own real SVG, not a redrawn approximation -- composited into a grid sized to whatever the day's round count is (roughly square, so a 7-game day and a 13-game day both lay out sensibly rather than one long strip). Games you answered correctly get a solid green border, wrong answers red, and anything you never got to before time ran out (skipped, or reached only after the clock hit zero) gets a faded, dashed border instead of a real color -- the same visual language the recap list already uses for an unscored attempt, so a not-yet-graded board never looks like a real verdict.

Results sync to your account the same way puzzle completions and ratings do: the client reports what happened, the server holds the actual history.