Mahjong, by LLMs alone — except this time only one of them finished. Same as the RowClear entry, only the harness differs: the hare gets one prompt with nothing in it but the idea; the tortoise gets a reviewed pipeline. Play the tortoise above — it's live. The hare has nothing to show, and that's the story.
The Hare Overloaded
On Bridge, the lazy one-shot strained past a clean message and squeezed out a whole game — barely. Mahjong is bigger. A real riichi game is a 136-tile wall, a dead wall and dora, draw-and-discard with chi/pon/kan calls, tenpai and riichi, and a han/fu scoring table that trips up humans. Ask for all of it in one prompt and the model doesn't strain to the finish — it fills the context and stops, holding a half-built thing you can't play. No file. No hare.
It's tempting to read that as "the one-shot approach failed — so much for the hare." That's the wrong lesson.
The Bet Wasn't Wrong — the Prompt Was Lazy
The whole point of the hare is a bet worth keeping: how far can an idea carry itself when you don't pollute it with process? No backlog, no tickets, no six-phase workflow — just the idea, stated once, standing up on its own. On small, sharp ideas that bet wins outright, and the tortoise's machinery reads as overhead. RowClear proved it.
Mahjong only found the ceiling of the laziest possible hare — one prompt, with no thought given to holding an idea bigger than one message. That ceiling is real, but it isn't the bet's. Throwing out "stand up an idea without ceremony" because a single naive prompt ran out of room is like deciding you can't carry groceries because you tried to hold them all in one hand. There are other quick ways to carry more — still unpolluted, still hare-spirited, just not lazy. We haven't tried them here yet. (That's a post for another day.)
The Tortoise, Meanwhile, Is Playable
The pipeline never needed the idea to fit in one message, because it never tried to hold it in one place. Grown ticket by ticket, the tortoise is a pure riichi engine — wall, deal, calls, waits, yaku, han/fu scoring — split from its view, covered by tests, then compiled to a single offline HTML file with no server. It's the box above. Open it and play a hand against three bots.
The Scoreboard, So Far
The tortoise is the only contestant standing — not because the hare's idea was wrong, but because the laziest version of it can't hold a game this size. So the open question the series leaves isn't "tortoise or hare." It's whether a smarter hare — quick, unpolluted, but wiser about scale — could stand this up too. Check back.