Sparse attention, one head at a time.

Tile shapes read as (temporal, height, width) in tile units. All six cover 128 keys, but they split that budget along different axes — so which one a head picks says something about where that head's important interactions live in the latent grid. Heads converge on different shapes as depth increases, which is how the packer hits its keep-ratio target without running a per-layer search.

The modeled speedup is a model, not a measurement. It is computed in-browser from the plan table: dense cost is ((t·h·w)/128)² per head; sparse cost sums n_tiles × (keep_ratio × n_tiles) over each head's chosen shape. It reproduces the trend — larger gains on longer clips, because dense cost is quadratic in latent tokens while sparse cost is near-linear — but it overstates the real figures (predicts ~7.8× at 5.17 s where the model card measures 4.75×) because it counts arithmetic only, ignoring kernel launch overhead, tile-boundary padding waste, gather cost, and memory traffic. The measured table in the README is authoritative.

plan_mse is regression error against the teacher's own head selection over 400 entries per geometry — lower is a closer fit to the teacher, not a measure of output quality.