A CONCEPT · PROJECTED FROM THE VINDEX KNOWLEDGE GRAPH
ATTENTION OUTPUT GATE
a primitive earned by refusal
A sigmoid gate on the aggregated attention output, multiplied in before the output projection — a generic primitive of the vocabulary, not a family quirk.
It entered the format by refusal: the first real four-norm model shipped a gate weight in all 52 layers, and the closure gate refused every one, naming the missing primitive. Its semantics were judged from the reference — sigmoid of the gate projection of the attention input, scaling the head output before o_proj — the primitive entered the IR, and the model closed at twelve of twelve operands per layer.
IN THE GRAPH
attention output gate— belongs to →attention
attention output gate— discovered by →operand closure