Method implementations

PLLM's MPCache-inspired work lives in crates/pllm-models/src/cache.rs. It is exposed through the pllm/kv-cache-eviction capability and Python KvCacheEviction component. Applying it to a dense Qwen2 or Qwen3 ModelPlan preserves fixed-capacity Key/Value state, adds explicit static-selection index state, and rewires decode attention through bounded dynamic gathers. It does not reproduce the paper's protected three-party runtime.

Use the component guide and generated component inventory for public parameters and capabilities. The MPCache paper page records provenance and limitations. Fidelity, protected execution, assurance, and matched benchmarks have not been evaluated for this structural adaptation.

A method that resembles a paper is not automatically a faithful reproduction. MethodRecord keeps classification, semantic scope, source/artifact links, native ownership, implementation level, assurance obligations, and five lifecycle statuses separate. ResearchRegistry.promotion() permits a native operator/region/full model only when full text and artifact locks are resolved, upstream reproduction evidence is recorded, fidelity/benchmark/promotion pass, protected execution and assurance pass or are explicitly not applicable, and at least one profile is eligible. Structural adaptations and planned/reference-only methods remain blocked.

Protected Q7 SiLU(gate) * up follows the same rule. Python exposes BinaryTableGatedMultiplyQ7 and R03CrtGatedMultiplyQ7 as interchangeable method components. ScalarProtectedTensorSchedule and IndependentLanesProtectedTensorSchedule select scheduling separately. All four are distinct component identities in two capability families; R03 is provenance for one implementation, not a package boundary. Four lanes are the current maximum, and no selection establishes complete-decoder coverage or cryptographic assurance.