05 / PORTABLE CACHE

Carry the work.
Keep its meaning.

Portable cognitive caches give reusable computation an explicit identity. PD-native connects those artifacts to a language for describing workloads, dependencies and compute intent.

01

Reuse begins with identity.

Content-addressed blocks carry sparse vectors or exact key/value state. Import checks namespace, conditioning, generation, shape, finite values, checksum and contiguous parent history before admitting a block. Matching bytes alone do not establish a compatible workload.

02

Freeze before you carry.

The cache checks the resident-to-carried transition with symbolic codegen (PandoraSymbolicLinalgLib), Salsa and Z3 before export. Only frozen exports travel; live step state stays local. Invalidated records cannot be revived, and a new generation requires a new record. Hosts must authorize tenant transfers: a checksum verifies content, not the sender.

03

Describe intent in PD-native.

The version-2 .pd grammar gives symbolic workloads a typed path through pd-native check and plan. Validated formulas lower into compiler source artifacts and launch plans. The broader language direction brings portable caches and compute objectives into the same declaration; a source-artifact plan does not execute a GPU kernel.

04

A portable contract, with explicit boundaries.

The cache crate provides bounded blocks and a validated JSON transfer envelope. Tensor conversion and network transport still need runtime integration. Compiler memoization is local and separate from portable cognitive state; portability does not make compiled GPU artifacts interchangeable across hardware.

REPOSITORY MAP
crates/dk-cognitive-cache/README.md · pandora-codegen/README.md · PANDORA_PD_LANGUAGE_PLAN.md

THINK IN POSSIBILITIES. BUILD WITH CONSTRAINTS.

A better place for
your next big idea.

Build with Pandora