A closed algorithmic system cannot be audited from inside a rented walled garden. Cloud-based AI inference carries inherent risks for proprietary data ingestion: hidden caching, swap-file leakage, and the fundamental absence of a verifiable chain of custody between the operator and the model. None of these risks exist on bare-metal hardware under direct physical control.

The value of this node is not the components. It is the topology. The architecture enforces a Bimodal Security Constraint: for adversarial audits and proprietary ingestion, Mode 0 applies - a complete physical air-gap. When broad external extraction is required, Mode 1 applies - frontier models are engaged strictly as untrusted sub-contractors. Proprietary metadata is stripped before transmission via the SEGA protocol. The returned payload is treated as contaminated until it passes the local Independent Protection Layer. The cloud is never trusted. The local interlock is.

This architecture is hardware-agnostic. The ISN and IPL components can operate locally while the AG is accessed via a remote API. The pattern applies to cloud-deployed frontier models without requiring cooperation from the model provider. The local interlock retains its authority regardless of where inference runs.

Specifications provided as reference for replication. Not a commercial requirement. The architectural pattern is what transfers.

The Epistemic Control Architecture runs three inference nodes sequentially. Every query passes through all three before the operator receives any output. No token is rendered until the IPL issues its verdict.

The pipeline enforces deterministic trip states. These are not error messages. They are non-repudiable audit records of exactly why the system halted. [ILK-DEV] and [ILK-BYP] are the primary epistemic interlocks. Both are quarantine states requiring explicit logged operator release action.