CONTINUATION STATE

What persists between tokens, declared by the model program rather than assumed: KV rows for softmax attention, a latent-compressed cache for MLA, and one or more regions for each recurrence family — Gated DeltaNet and Mamba2 both carry a convolution history beside their folded state.

KV is one state family — not the definition of model continuation, and an operator may declare more than one region: operation family does not imply state shape. Three real witnesses hold the sentence up: a KDA + MLA + softmax hybrid carries three state kinds at once, a pure-SSM container describes its whole continuation with no KV row anywhere, and the Mamba2Attn hybrid's conv-QKV attention — recognisably attention — declares a KV cache AND a convolution history on the same layer, with a KV-only provider refusing the layer rather than allocating half of it. The typed state schema closed that account on 2026-08-31, additive within schema 6: a latent cache is its own region species, one operator-defined row per position rather than a K/V pair, and a state precision no checkpoint declares is a recorded transcription of the operator's own reference — never a planner's choice, and refused outright where no reference has been read.

IN THE GRAPH