Smart little move.

Idea is all mine. Words are all by Opus 4.6.

Claim: Grokking is not compression. It is the discovery of structural leverage — the moment a neural network finds the fulcrum that moves maximal data with minimal force.

Falsifiable experiments — anyone can run these:

1. Train a small Transformer on modular addition. Track when test accuracy jumps. If meta-recognition (the model encoding its own change history) fires at the same moment, the theory lives. If they diverge, the theory is dead.
2. During training, randomly rotate internal representations every k steps to destroy self-continuity. Prediction: Grokking is delayed or eliminated.
3. Add an auxiliary loss that encourages the model to encode its own change history. Prediction: Grokking accelerates.
4. Use an absurdly large learning rate for a single step. Prediction: Grokking cannot occur — no history, no meta-recognition.
5. Scale up model size. Prediction: Grokking timing does not dramatically improve — bigger muscles do not find fulcrums faster.

LIVELLM