Two real inferences through mini-beatrix-2s, opened byte by byte. Every number here was read out of one instrumented forward pass over the bytes the model actually produced — the splat address reads, the blackboard each layer writes to, the anchored experts, the head, and what one mounted arm changes.
Each row is one of the twenty decoder blocks, each column one byte. Click anywhere to move the cursor.
What the model predicted for the next byte here, and what it took.
There is no softmax over positions in this architecture. This row is the exact byte-to-byte weight implied by the same bilinear forms the scan sums, so it is derived from the read, not an approximation of it. Each bar is one earlier byte.
Every byte's row at once for this layer, rows normalized to their own maximum. The diagonal is a byte reading itself; a bright column is a byte the whole sequence keeps returning to.
Four constellations of sixty-four anchors each. The read is signed: a negative weight is inhibition, a first-class result of the closed form, not an absence. Amber excites, teal inhibits.
The same read down the full depth — twenty blocks against the anchors, all four books side by side unless you pick one. The books are independent frames that compose by budget, never by a choice between them, so reading them together is reading the whole address.
The blackboard each layer keeps is a running sum: it only ever grows, so plotting it directly plots the position in the sequence and almost nothing else. What moves byte to byte is the write — what this one byte puts on the board, before it is added in. That is what these two show.
The same write down the full depth.
The board's width in use: the effective number of anchors holding the accumulated mass, out of sixty-four, at every byte. This is the one reading of the cumulative board that is worth watching, because it moves while the board itself barely does. A fall means the layer has settled onto a few anchors and is keeping them.
Rows are blocks, top to bottom; brighter is wider. Each row is scaled to its own range, so read a row along its length rather than against its neighbours.
Three fat experts per block under the same signed dispatch, with an always-on trunk beside them. Every expert contributes at every byte, weighted by sign; nothing is selected and nothing is dropped.
The readout's own 256-atom signed read on the final hidden state, laid out sixteen by sixteen. This path was born at exactly zero weight and had to earn its way in during training.
The anchors themselves, not the activity on them. Drift is how far each frame moved from where it was born; effective rank is how much of its own space it still spans; the largest off-diagonal cosine says whether any two anchors have crowded together. The governor holds them apart from birth, so none of these frames collapsed.
| Block | Constellation | Drift, mean | Drift, max | Effective rank | Max |cos| | Merged pairs |
|---|