GLASSBOX
glassbox
An AI agent is dropped into a game
it has never seen.
no name · no goal · no idea what any symbol means
▾
the usual way
Normally you'd train an agent
over millions of rounds until it wins.
millions of games
of trial and error
→
a neural network
that wins
→
ask it
why
The rules stay locked inside the weights.It can beat the game, but it can't explain it.
▾
why we built this
What if you wanted to
reverse-engineer the game instead?
Recover the rules themselves, as runnable code.
A game designer could point this at their own game and find the logic flaws before players do.
A modder could recover mechanics nobody documented.
And anyone can read exactly what the agent believes:
actual output: code the agent wrote for one of the games below
▾
how it works
It proposes rules and writes them as code,
then tests that code against everything it has seen.
play & observeevery move logged
→
propose ruleswritten as code
→
replay reality through the codeevery cell must match
⟲
wrong? fix itthe failure goes back in
reality
what its code predicted
One wrong cell, like the ringed one, refutes a proposal.It revises until nothing reality deals can break it.
▾
see it work
Start with games you know.
learned from a blank slate, then decompiled back to code
▾
the real test
But those games are in every training set.
So we invented new ones.
Same method, same agent, on rules we made up hours before the run.
No training data can help.
Our invented rules are stated on each card. The agent had to find them the hard way.