Read the explanation
The original page is an interactive display of stored attention-like lookup tables, not a transformer inference engine. Its third native sentence contains seven tokens, while its first contains eleven. Choosing a sentence changes how many query rows and key columns the native matrix renders. Here every held square corresponds to one stored entry. Seven times seven gives forty nine entries; eleven times eleven gives one hundred twenty one. The grid changes discretely when the actual sentence option changes. These figures do not invent a fourteen token option, fractional token counts, or a computed value-vector output. Counting entries illustrates why a dense square table grows with the square of its token count. It does not prove the runtime of a trained transformer implementation. The original native tables contain some rows whose numbers do not sum to one, so this explainer will label them as assigned weights rather than silently normalize the product source or claim every table is a probability distribution. Keep the third native sentence and the coref head selected. Choosing tokens as the query selects its exact seven-entry stored row: zero point one, zero point four, zero point two, zero point zero five, zero point one five, zero point zero five, and zero point zero five. Choosing process instead selects zero point three, zero point two, zero point three, zero point zero five, zero point one, zero point zero two, and zero point zero three. Both rows happen to sum to one, and every key token has an individual value label. Both held cases use four hundred twenty pixels per assigned weight unit. The process key drops from zero point four to zero point two when the query changes, so its bar width drops from one hundred sixty eight to eighty four pixels. The source token buttons and matrix cells both select a query row; clicking a matrix column does not compute a new key score. This is a causal comparison of the actual native query action. It does not compute dot products, softmax, or value-vector mixing, and these authored weights do not establish learned grammatical relationships. Hold the third native sentence and tokens query. Change the actual head selector from coref to context. The context row is zero point one five, zero point three five, zero point one five, zero point zero five, zero point two, zero point zero five, and zero point zero five. Compare it with the full labeled coref row. The process key changes from zero point four to zero point three five, while parallel changes from zero point one five to zero point two. On the same four hundred twenty pixels per assigned weight unit scale, those parallel bars are sixty three and eighty four pixels. Both selected rows sum to one, but that statement applies to these two actual rows, not every table in the source. The native inspector sorts these assigned weights, and its arcs format their widths from the same lookup values. Head names are teaching labels. This page does not train heads or produce a semantic prediction, so the comparison explains the stored table selection without claiming measured learned behavior or a computed transformer output.