A CONCEPT · PROJECTED FROM THE VINDEX KNOWLEDGE GRAPH
QUERY SCALE vs SCORE SCALE
two multiplies, two addresses
Query and score scales are separate operations: the declared factor multiplies the normalised query before position encoding; the canonical head_dim^-0.5 is a score-time multiply.
Folding them into one multiply is algebra-equivalent but not fp-equivalent — moving a multiply across RoPE and a matmul is exactly the silent normalisation that later surfaces as a parity mystery. So the surface keeps them unfolded: query_scale, score_scale, and logit softcapping, each applied at its own declared point.
IN THE GRAPH
query scale vs score scale— belongs to →execution surface
query scale vs score scale— belongs to →attention