ORIGINAL TRAINING
PATCH'S PRACTICE TABLE
A little practice.
A little more room.
Two trained models, the same tricky quilt. Watch them choose a home for every patch—and see why leaving a little room matters.
No countdown. Nothing to win here. Just a little curiosity.
The trained models are loading.
AFTER PRACTICE
A little more foresight
These six quilts were chosen to show the original model getting stuck and the practice model finishing. They are illustrations; the separate test results are below. Your nudges can change the live result.
A peek inside these little brains actual inputs, values & tree paths
Choose any legal move to inspect its learned value. Inspection and previews leave both quilts where they are. Values are scores, not chances of winning.
| Visible input | Original | Practice |
|---|
The original forest uses 14 inputs. Practice adds eight geometric packing bounds, including whether enough separate spaces remain for larger patches. Exact finishability is an offline training label; neither model consults that search while playing.
What did the training actually improve?
| Policy | Fresh test quilts completed |
|---|
Models were trained in Python with pytoolz/WSRF. Watching them here evaluates the exported trees; it does not retrain them. The existing game's model remains unchanged.
Read the full training record ↗ · Seeds, parameters & results