Day 109 — the arcade, and games as the gauntlet

“i think we might be able to beat the 30B”

The coding axis is already 1:1. Today made the case that the general axis — the wide, per-clause, owned-only gauntlet — could be a fair fight too. Not with a bigger model: with a growable graded library, a router that knows where the cards live, an honest tongue, and games as the general-reasoning gauntlet.

The arcade

What began as a training roadmap became a full arcade. First the headless engines — Risk (exact combat odds), Catan (the exact 2d6 production distribution), Nim (the smallest provably-optimal game), a murder-mystery deduction engine — then their playable tiles, and a Murder Mystery that became the living Werbos Commons: NPCs wander a top-down pixel map and you click the cast to interrogate them.

Then the flood: a per-game Rules button and category grouping, Solitaire / Minesweeper / Ticket to Ride, Go Fish, casino Baccarat & Roulette & Video Poker & Craps, Dominoes + Mexican Train, Pac-Man, Streets (a GTA2-style sandbox with a fixed 64×64 core and procedurally-generated edges), a Doom raycaster, a Platformer, Crazy Eights, and Cards Against Humanity + What Do You Meme judged by werbos’s own tongue.

Then 26 more in a single pass — Gomoku, Dots & Boxes, Mancala, Nine Men’s Morris, Hex, Backgammon, Yahtzee, Farkle, Spades, President, Euchre, Gin Rummy, Cribbage, No Thanks!, Old Maid, War, Sushi Go!, Speed, Skull, Love Letter, Werewolf, Scattergories, Boggle, Snakes & Ladders, Ludo, Chinese Checkers. Each with its own AI, a 2–6 player selector where it fits, and its own visual identity. All typecheck-clean and deployed to the live box.

Tarot is the thesis in a costume

cards → tongue: the owned model reads the drawn cards and speaks them. It immediately mis-routed — “tarot card” fell into the code lane, “excess” into the teacher lane — so it grew an honest grounded fallback and a lane-leak guard that refuses to bluff. A game that doubles as a test of our own lookups.

Games are the gauntlet

The real turn of the day: a deterministic game is a reasoning-problem oracle with a non-model adjudicator. Nim proved the pipe — its best_move oracle now both generates and grades frozen ratchet probes (derive the optimal move, or honestly refuse from a lost position), 26/26 green. It ran beside a standalone reasoning ruler that climbed 4 → 18/18 across the day (derive 14/14, honesty 4/4): irregular-plural syllogisms, modus ponens, a compare-lane cardinal sin killed on the way. Chess/Reversi/Nim are hard deterministic tests; Wordle/Minesweeper clean puzzles; Tarot the honesty boundary. The general axis of the 30B fight, fuzzed by play.

The mint learned to think less, and to go deep

--new-domains was returning empty: Qwen3 ignores /no_think in the prompt and burns its whole budget in <think>. The fix disables thinking through the chat-template flag for real. Then --focus biases new-domains toward a slice of the map, so two rented boxes can split the knowledge space without overlap.

And with the graders wrapped (201k keepers of 205k staged), a rented Qwen now fires a thinnest-first depth burst: 16-way, always firing, the worklist sorted by source + staged size so it walks UP the ladder — raising the median from ~96 toward 500–1000 across the whole shallow tail instead of re-hammering the same domains. ~26k cards/hour; a 16-hour run adds ~440k cards (≈2.2× the 368k graded corpus), and ~$150 clears the 1M-card goal. Everything staged and ungraded, gated by an independent OLMo pass before it touches the live library.

The thought that stuck

One $200 shopping spree over 24 hours bought more data than months of minting ourselves and getting rate-limited. Scale the spend, scale the win — and now there’s a general-reasoning gauntlet to prove it on. The number decides, not the vibe. But the vibe is earned.