FREE · WINDOWS · MACOS · LINUX · MADE IN GODOT 4

TRAIN A MIND.
ENTER THE GRID.

Learn real reinforcement-learning techniques by playing a game. You don't code the fighter — you move nine sliders that say what it gets paid for, and a real search spends thousands of simulated matches finding a mind that maximises exactly that. Including the parts you didn't mean.

Play free on itch.io Watch the 42-second trailer

01 — THE LOOP

Three moves.
No code anywhere.

01
Design the reward

Nine weights — win, hit, taken, time, survive, accuracy, aggression, ricochet, flank. That is your entire input. You never touch behaviour.
02
Choose what it can see

Sixteen senses to pick from: aim error, range, incoming disc, time to impact, line of sight. Switch one off and it is blind to that fact forever — no amount of training brings it back.
03
Train, then fight

Press TRAIN and watch the learning curve climb live. When you like it, SAVE & FIGHT drops the trained policy straight into the Grid against a champion.

02 — THE CATCH

It learns exactly
what you pay for.

Not what you meant. What you wrote. Every time.

You pay for SURVIVAL

It hides behind a pillar and never fires. Scores brilliantly. Loses every single fight — a flawless coward.

You punish TIME too hard

It walks out of the ring to end the round early. It solved the problem you actually gave it.

Closing the gap between the reward you wrote and the outcome you wanted is reinforcement learning. The coach panel watches your two curves and calls it out the moment they diverge.

03 — HONEST MACHINERY

This is not PPO.
It's CEM.

The cross-entropy method: sample a population of whole minds, score each over simulated matches, keep the elite few, re-aim at those, narrow as the budget runs down. No gradients. No backpropagation.

Which is why you will not find a learning rate or a discount factor anywhere in this game — this method does not have them, and inventing knobs that do nothing is how tutorials teach people wrong. In advanced mode you drive the search yourself: population, elite fraction, exploration, generations.

0: reward weights
0: senses to pick from
0: minds per generation
0: champions to beat

04 — WHY WE BUILT IT

Play is the future
of learning.

Ask an AI, and it hands you the answer in four seconds. What it cannot hand you is the ten minutes you spent training a coward—watching a machine do exactly what you asked, and lose anyway.

Nobody remembers the textbook definition of reward hacking. Everybody remembers their own fighter walking out of the ring to stop losing points.

That is the bet MINDFORGE is built on. Explanations are free and everywhere now; what is scarce is the experience of being wrong yourself, on purpose, with something you built. So MINDFORGE is a game first—but underneath, it runs on a real cross-entropy search, real reward functions, and real failure modes. Because a lesson you cannot reproduce isn't a lesson at all.

05 — THE GRID

Then it proves it.

Your trained policy plays the exact behaviour it learned — no scripting, no cheating. Three cameras, a contracting ring, best of five. Or take the disc yourself and find out how hard your own reward really was.