FROM DESCRIPTION TO COMPUTATION

SURFACES FOLLOW THE PROGRAM — SCHEMA 6, WITNESSED

DecoderStack
      ↓
attention assumed
      ↓
FFN assumed
      ↓
KV assumed

Component
      ↓
declared operator program
  softmax · MLA · KDA
  gated-delta · mamba2
  conv-QKV attention · …
      ↓
required surfaces
      ↓
continuation state

mamba2-780m — 48 declared Mamba2 operators · 0 attention operators

before schema 6:   48 fabricated attention surfaces   ✕
at schema 6:       48 Mamba2 operators
                    0 attention surfaces
                    0 FFN surfaces                     ✓

mamba2attn-250m — 28 Mamba2 operators · 4 conv-QKV attention operators

at schema 6:       attention exactly where declared
                   conv-QKV, never plain softmax
                    0 FFN surfaces                     ✓

Both directions of the sentence are now executed evidence, not schema tests. An architecture with no attention fabricates none: the pure-SSM witness was first refused with nineteen itemised findings, then admitted with zero fabricated surfaces and operand closure enforced at encode. And an architecture with selected, non-standard attention gets those operators only where declared: the hybrid witness carries its four conv-QKV blocks — causal convolution over the fused QKV, partial rotary — as their own operator on exactly the declared layers, and runs them at its reference's numerical floor. The Record keeps every half.

OPERAND ROLES — THE TYPED VOCABULARY

Every tensor inside a decoder stack is typed by what it is to the generic operations. An operation plan binds every operand its programme requires, or refuses. Closure needs the complete set — a missing operand is a named refusal, never a silently skipped step.

AttnQ  AttnK  AttnV  AttnO
AttnQBias  AttnKBias  AttnVBias  AttnOBias
AttnOutputGate  AttnSinks
AttnQNorm  AttnKNorm

LinearAttnInProjQkv
LinearAttnInProjA  LinearAttnInProjB
LinearAttnInProjZ  LinearAttnConv1d
LinearAttnALog  LinearAttnDtBias
LinearAttnNorm  LinearAttnOutProj

Mamba2InProj  Mamba2Conv1d
Mamba2Conv1dBias
Mamba2ALog  Mamba2D  Mamba2DtBias
Mamba2GatedNorm  Mamba2OutProj
Mamba2PreMixerNorm

ConvQkvInProj  ConvQkvConv1d
ConvQkvConv1dBias
ConvQkvOutProj
Mamba2PreMixerNorm

KdaQProj  KdaKProj  KdaVProj
KdaQConv1d  KdaKConv1d  KdaVConv1d
KdaFAProj  KdaFBProj
KdaGAProj  KdaGBProj  KdaBProj
KdaALog  KdaDtBias
KdaONorm  KdaOutProj

MlaQProj  MlaKvAProj  MlaKvBProj
MlaKvANorm  MlaOutProj

FfnGate  FfnUp  FfnDown

PreAttentionNorm  PostAttentionNorm
PreFfnNorm  PostFfnNorm
PreExpertsNorm  PostDenseFfnNorm
PostExpertsNorm
placement: PreOnly | PrePost | PreMixer

MoeRouterWeight  MoeRouterBias
MoeRouterScale  MoeRouterPerExpertScale
ExpertGateUp(+Scales/Bias)
ExpertDown(+Scales/Bias)
PerExpertGate/Up/Down(u16)
SharedExpertGate/Up/Down
LayerScalar

THE COMPILER BOUNDARY — WHERE FAMILY KNOWLEDGE IS ALLOWED TO LIVE

HF model_type, upstream reference
        ↓
architecture-specific judgment OK
        ↓
generic VINDEX3 IR

VINDEX3 IR
        ↓
generic op plan
        ↓
generic kernels
— no family knowledge here

VINDEX3 permits architecture-aware ingestion and produces an architecture-independent executable IR. The trap this guards against has a name — the naming-convention trap. Dispatching on object-id strings is the same defect laundered through a name: the convention becomes an undeclared schema.

STATE IS AN OPERATION REQUIREMENT

Operator
   ↓ declares
ContinuationState — one or more typed regions

softmax attention    └─ KV rows              grows with context
MLA                  └─ latent KV            compressed, per position
gated deltanet       ├─ delta matrix         fixed size, folded every token
                     └─ conv history
Mamba2               ├─ SSM state            head_dim × state_size per head
                     └─ conv history
conv-QKV attention   ├─ KV rows              grows with context
                     └─ conv history         fixed, over the fused QKV
future operators     declared, never assumed

An operator may declare more than one continuation region — and operation family does not imply state shape. Both recurrence families here carry a convolution history beside their folded state, and modelling only the matrix once left a whole buffer invisible until the first single-token continuation. Nor is multi-region state peculiar to recurrence: the hybrid witness's conv-QKV attention — recognisably attention — requires a KV cache AND a convolution history on the same layer, and a provider that can hold only rows refuses the layer rather than allocating half of it. KV is one region kind — not the definition of model continuation. Three real witnesses hold that sentence up: a KDA + MLA + softmax hybrid carrying three state kinds already executes, a pure-SSM container describes its whole continuation as two recurrent regions per layer — eighteen million elements, constant in sequence length, no KV row anywhere — and the Mamba2Attn hybrid runs both mixed regions through generic execution at its reference's own numerical floor. The typed state schema — declared precision for KDA, latent-cache geometry for MLA — closed that half of the lift 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 Kimi-Linear-48B now declares its 20 recurrent layers and its 7 growing caches from the container alone. The rule underneath does not change: state geometry is a container fact, read from the plan, never inferred from architecture — and the surfaces that report it read the same plan, after one summary was caught calling seven growing caches constant-size recurrent state.