猜規則

是，而且是寫在程式裡的。每一局，系統會先在真規則「裡面」找一條更窄的規則，兩張範例都從那條窄規則裡挑。用經典的 2-4-6 來打比方：如果真規則是「由小到大」，範例卻剛好是 2、4、6——它們同時也符合「每次加 2」這條窄很多的規則。窄規則整個包在真規則裡，所以不管你照著窄規則丟幾張卡，盒子都會回答符合。這不是在耍你，這就是整個遊戲要讓你親手撞上的東西。揭曉時畫面會直接點名你今天這一題掉進去的那條窄規則。

1960 年英國心理學家 Peter Wason 做的實驗。他跟受試者說：「2、4、6 這組數字符合我心裡的一條規則，你可以自己提出三個數字來問我符不符合，覺得懂了再說答案。」多數人很快想到「每次加 2」，然後問 8、10、12，問 20、22、24，每次都得到「符合」，於是很有把握地宣布答案——錯了。真正的規則只是「由小到大」。問題不在他們不聰明，而在他們問的每一組都是「我的猜測會說符合」的數字；那種問題永遠不可能證明你錯。真正有用的問題是 1、2、3 或 6、4、2 這種「如果我是對的，它應該不符合」的組合。這個遊戲就是把那個實驗搬到卡片上，再加上一件原實驗沒有記錄的事：你每一次押的是哪一邊。

確認偏誤是傾向去找「支持自己想法」的證據，而不是去找「能推翻自己想法」的證據。麻煩的是它從外面很難看出來：兩個人問的問題數量可能一樣，拿到的「不符合」次數也差不多（模擬裡只差 0.2 對 1.6 次），真正不一樣的是他們預期會發生什麼。所以這裡的投遞口就是押注——你選左邊，就是在說「我猜會符合」。揭曉時攤開的押注紀錄，像是「押『會符合』9 次、押『不會符合』0 次」，才是你這一局的測試策略的實際樣子。它量的是這一局，不是你這個人。

你押的和盒子回的不一樣，就叫意外：你把卡從「我猜不會符合」丟進去，它卻落在符合那一疊。卡片會從你投的那一側橫越到另一側、翻成白底、打上驚嘆號。會這樣設計，是因為一張照你預期落下的卡，只是讓你原本的想法多撐一次；一張讓你意外的卡，才真的刪掉了一個錯的想法。科學家做實驗要找的就是這種結果。分享卡上的 ❗ 越多、而且越早出現，通常破解得越快。

考卷沒過，錯的卡會翻開真答案，八張全部放進證據堆，次數加上「8 ＋ 答錯張數」。會是 8，是因為交卷那一刻你等於看到了八張卡的答案，而且那八張是專門挑來戳破錯誤想法的，比你自己隨手丟的卡更有資訊。我們用模擬跑過：如果只罰 3 次，一個完全不測試、直接交卷然後靠考卷答案猜規則的玩家，會在 99.7% 的題目上比認真測試的人更省；改成 8 加上答錯張數之後，這種玩法在同一題上贏的比例是 0%。差一張就錯的人罰得比亂猜的人輕，這也是故意的。

用你所在地的日期當種子，同一天全世界拿到同一條規則、同一組範例、同一份考卷，連續兩天一定不會是同一條。分享出去的只有日期、一串 ⚪⚫❗ 和次數，不會出現規則的文字，可以放心貼到群組。自由練習則是另外抽題，可以選難度一到三，不影響連續紀錄。

Yes, and it's built into the code. Every puzzle, the engine first finds a rule strictly narrower than the real one, and draws both examples from it. Using the classic 2-4-6 as an analogy: if the real rule were "ascending order", the examples might happen to be exactly 2, 4, 6 — which also satisfy the much narrower "add 2 each time". Since the narrow rule sits entirely inside the real one, any card that fits it always answers "fits" too. That's not a trick on you personally — it's the exact thing the game wants you to run into. The reveal names the narrow rule today's puzzle drew you toward.

A 1960 experiment by British psychologist Peter Wason. Subjects were told the sequence 2, 4, 6 fit a rule in his head, and could propose their own triples to test before announcing their guess. Most people quickly guessed "add 2 each time", tested 8-10-12 and 20-22-24, got "fits" every time, and confidently announced their answer — wrong. The real rule was just "ascending order". The problem wasn't intelligence; it was that every test they ran was one their own guess predicted would fit, and that kind of test can never prove you wrong. A useful test is one your own guess predicts should FAIL, like 1-2-3 or 6-4-2. This game puts that experiment on cards, plus something the original never recorded: which side you bet on, every single time.

Confirmation bias is the tendency to look for evidence that supports what you already believe, rather than evidence that could disprove it. It's hard to see from the outside: two players can run the same number of tests and get nearly the same number of "doesn't fit" answers back (0.2 vs 1.6 per game in simulation) — what actually differs is what they expected going in. That's why posting through a slot IS the bet. The reveal's ledger — "bet fits 9 times, bet doesn't fit 0 times" — is what this run's testing strategy actually looked like, in your own numbers. It measures the run, not you.

A surprise is when your bet doesn't match the box's answer — you post through "won't fit" and it lands in the fits pile anyway. The card visibly crosses to the other side, flips to a bone-white face, and gets a heavy exclamation mark. It's built that way because a card that lands where you expected only reinforces a theory you already had; one that surprises you actually rules a wrong theory out. That's exactly what a real experiment is for. More ❗ marks, especially early, usually means a faster crack.

A failed exam flips the wrong cards to their true answer and adds all eight to your evidence, and the count goes up by "8 plus the number you got wrong". It's 8 because the moment you submit, you've effectively seen eight answers — and they were chosen specifically to expose the wrong theories a player would likely hold, which makes them more informative than cards you'd post yourself. We simulated the alternatives: a flat +3 penalty lets a player who never tests, just declares immediately and mines the exam's answers, beat an honest tester on 99.7% of puzzles; 8 plus wrong-count brings that down to 0%. Missing one card also costs less than a wild guess — that asymmetry is intentional too.

Your local calendar date is the seed, so everyone in the world gets the same rule, the same examples and the same exam on the same day, and it never repeats on consecutive days. Sharing only sends the date, a string of ⚪⚫❗ symbols, and a count — never the rule's text, so it's safe to post in a group chat. Practice mode draws a separate puzzle at a difficulty tier of your choice and never touches the daily streak.