A CONCEPT · PROJECTED FROM THE VINDEX KNOWLEDGE GRAPH
HYBRID ATTENTION POLICY
each layer declares its family
A stack that mixes attention families — delta then full, sliding then global — declared layer by layer in the persisted policy table, never computed from a layer index.
The per-layer AttentionLayerPolicy is the single authority: span, window, position, family. No layer-modulo arithmetic exists on the execution path — the interleave is unrolled into the table at compile time, and mutating one row provably changes execution. That is how a 3:1 delta-to-full stack executes with zero architecture branches.
IN THE GRAPH
hybrid attention policy— read by →execution surface
hybrid attention policy— governs →kda — kimi delta attention
hybrid attention policy— governs →attention