Read the explanation
This tool generates two reproducible synthetic four-component query, key and value heads from token positions. The vectors are not trained embeddings, and editing a token word without changing its position does not learn its meaning. In the default sequence, choose the second token, Twitchs, in head one. Its query is point one five, negative point zero three, negative point nine one, negative point six four. The dot product with the first key is negative point zero one seven eight, while the dot product with the second key is negative point seven one zero four. Dividing by the square root of four divides both by two. The two negative bars share the same two hundred pixels per score unit: the larger score stays closer to zero. Causal attention at the second token admits only the first two positions. Future scores become negative infinity, so their exponentials are exactly zero. Subtracting the largest allowed score before exponentiation preserves the normalized weights and makes the computation stable. At temperature one, Caught receives about point five eight five seven and Twitchs about point four one four three. Each bar uses four hundred eighty pixels per probability unit. Reduce temperature to point two and the larger score becomes much more dominant. The first weight rises to about point eight four nine six; the second falls to about point one five zero four. The allowed bars still add to one. Bidirectional mode restores the remaining tokens before recomputing the normalization. Attention mixes the value vectors, not the query vectors. At temperature one, the first context component is point five eight five seven times negative point two one, plus point four one four three times negative point five three. This gives approximately negative point three four two six, displayed as negative point three four because this tool rounds context components to two decimal places. The line drawing uses a shared three hundred pixels per component scale. The two weighted contributions add head to tail to reach the context component. Switch heads or presets to recompute every dot product, attention row and context. Select rows through the arc graph or any heatmap cell, edit two through seven tokens as the actual handler permits, and export the full attention matrix and rounded context vectors. This is an inspectable synthetic demonstration of the algebra, not a language-model output or benchmark.