CASK | Cognition-Aware Sketch KV: A lane-conditioned, verifier-gated KV-cache substrate for the MABOS cognitive runtime

CASK, a Cognition-Aware Sketch KV is an independent research report and implementation specification for a MABOS-native KV-cache substrate designed for long-context local inference on workstation-class hardware. The paper introduces a lane-conditioned, verifier-gated memory hierarchy that replaces flat KV storage with provenance-aware tiers: exact protection for identity, operator intent, tool evidence, audit-sensitive spans, and other cognitive anchors; protected-subspace scoring for attention-critical directions; QJL residual sketches for compressed long-context state; and Dream-lane policy adaptation driven by Compression-Induced Divergence Rate against the Cold Lane auditor. Rather than treating KV cache as anonymous tensor storage, CASK frames it as a cognitive working-memory substrate whose compression policy is derived from runtime provenance, hardware constraints, and operational verifier feedback. The report includes mathematical derivations, error-budget analysis, hardware mapping for AMD 7950X3D / RX 7900 XTX, and acceptance-gated implementation steps for the MABOS cortex.os runtime.