Model anatomy lab

Orbit Q/K/V branches and residual paths. Calculate a small attention block and inspect a sampled neuron atlas.

How this experiment works

Layouts adapt Token-Print and Neuron Atlas. Sinusoidal positions, scaled dot products and stable softmax calculate real toy attention weights. The full matrices can be exported.

What it does not prove

The eight-dimensional block is untrained and uses whitespace pieces. Atlas dots are structural samples, not measured activations. Governance surrounds tools; it does not certify weights.

Try this challenge

Turn the causal mask off. Which future pieces gain weight? Then explain why a governance field is separate from the attention calculation.

Enable JavaScript to run the experiment on your device.