Used by checkpoint 460
Build log · checkpoint 460
How we taught a small model to see the board.
We turned a real game engine into a teacher, checked every move, and trained Qwen 3.5 4B on the hard moments—not just the common ones.
What “93.9%” means: 612 correct actions on our 652-case raw-state game suite. It is not a general intelligence score.
RECORD PRESSLIVE
RAW STATE24 · 25 · 26 · ?
LEGAL LABELBID $4
10K recordsRAW prompts4B model
10,000checked lessons
652held-out tests
93.9%correct actions
100%valid JSON
The simple version
A game becomes a lesson.
The engine makes a legal moment. The teacher chooses. The filters refuse anything untrustworthy.- 01DATA
Play
The real rules engine creates a legal state.
- 02DATA
Ask
The teacher prices every legal action.
- 03DATA
Check
Bad JSON, illegal moves, and contradictions are rejected.
- 04DATA
Balance
Rare wins, blocks, folds, and placements keep their seat.
- 05ADAPTER
Learn
Qwen sees the same compact raw prompt used in the app.
Used by checkpoint 460
Targeted hard cases
Wins, blocks, expensive folds, Payday choices, and Space 49.Ready for the next dataset
Live Convex games
The ingestion lane exists, but the selected manifest does not list it.Try the curriculum
We teach the moments random games miss.
Pick a case. The board changes, but the question stays small: bid, fold, or place.1
2
3
4
5
6
7
24
25
26
27
28
29
8
23
40
41
42
43
30
9
22
39
48
49
44
31
10
21
38
47
46
45
32
11
20
37
36
35
34
33
12
19
18
17
16
15
14
13
Your chip Rival chip Card space
WIN lesson
One card ends the game.
You own 24, 25, and 26. Space 27 makes four in a row, so the model must bid.
The model sees
OPEN · space 27 · minimum $4EXPECTED MOVE“bid” at $4
The trust layer
Bad lessons never reach the model.
A legal-looking sentence can still teach the opposite move. We check the structure, the game, and the words.- 1Replay the state
The production reducer must accept the action.
- 2Check the explanation
“Fold” cannot end with “raising to $8.”
- 3Remove conflicts
One visible prompt cannot teach two different actions.
- 4Protect the exam
Held-out prompts are deleted from training.
bid $4fold / “raising”place 77bid $999
GATE AValid JSON?
GATE BLegal move?
GATE CWords agree?
APPROVED
{ "action": "bid", "bidAmount": 4 }REJECT
The expensive surprise
Lower loss can hide a worse player.
Tap the checkpoints. The training number improves while the game decisions collapse.VALIDATION LOSS0.1422lower looks better
REAL GAME ACTIONS63%Better policy
task accuracy
The rule we kept: save early checkpoints and choose with the game exam—not the prettiest loss curve.
The honest scorecard
Excellent at auctions. Still learning where to place.
The overall number is real, but the buckets tell us what to improve next.93.9%
612 of 652 raw-state decisions
CHECKPOINT 460Why the gap?An auction usually points to one space. A Payday card can offer sixteen. Ranking many legal choices is the next data problem.
Small on purpose
A specialist carries less—and knows exactly where it is going.
A frontier model rents enormous general capacity. Our adapter spends its capacity on one board, one policy, and short JSON.GENERALIST
Frontier model
SPECIALIST
Qwen 4B + LoRA
FINISH
On its home task, the small model starts with the exact board language, rules, and action format. Specialization can beat raw scale.
VERIFIED TODAY93.9% internal raw-state suite
NEXT PROOFSame-suite frontier eval + seeded arena games
We explain why it can win here. We will say it has beaten a named frontier model only after the matching report exists.
From checkpoint to game
Cheap while idle. Fast after the first wake-up.
Scale-to-zero saves money, but cold starts are part of the product experience. We measured both.144.3sCOLD
6.5sWARM
Same prompt. Same 73-token answer. The difference is loading the model.
APPraw board
→MODAL L4checkpoint 460
→ENGINElegal action
T4 failed: no BF16 path, request crossed the 150-second web limit. Health alone did not catch it.
The next experiment is live play
Put the specialist at the table.
Choose “ZIL Qwen 3.5 4B” in a new match. Every real game gives us a better map of what the next dataset should teach.