Appendix C. Component Responsibilities and Interfaces

Table 11: Component Responsibilities

Component	Generic role	Reference implementation
TEE-capable CPU	VM-based trusted execution	AMD SEV-SNP or Intel TDX on a validated server platform
CC-capable GPU	GPU attestation, secure GPU session, GPU memory protection	NVIDIA GPU with CC mode enabled
Kubernetes platform	Admission, scheduling, RuntimeClass, RBAC, service discovery, network policy, and operations	Validated Kubernetes platform
Confidential runtime class	Selects the handler that launches the pod inside a measured confidential VM	Confidential containers runtime
Confidential guest	Attestation Agent, Confidential Data Hub, Kata Agent policy, model server, and secure GPU session	Kata-based confidential guest with Trustee guest components
Attestation verifier	Validates CPU, GPU, guest, image, runtime-policy, nonce, and policy evidence	Attestation Service
Key broker or KMS	Releases or unwraps model keys after attestation and policy evaluation	KBS with KMS/HSM integration
Encrypted model artifact store	Stores model artifacts without exposing unencrypted artifacts to the node or cluster	Registry, object store, or encrypted volume
Inference endpoint	Narrow HTTPS endpoint exposed by the confidential pod to approved callers	Kubernetes Service, route, gateway, or load balancer

Table 12: Component Interfaces

Interface	Producer	Consumer	Requirement	Required operator signal
Pod admission and scheduling	Kubernetes API, scheduler, RuntimeClass, node labels	Platform operator, runtime, verifier	RuntimeClass, node selector, GPU request, namespace, service account, image digest, policy annotations	Admission, scheduling, or runtime-class failure naming the unsupported setting
Confidential sandbox launch	Kata runtime and worker node	Platform operator, verifier	Guest image/initrd, runtime policy, CPU TEE flags, GPU assignment, pod identity, network devices	Sandbox launch success/failure naming runtime, node, guest, or device reason
Attestation evidence	Guest Attestation Agent and GPU attestation component	Attestation Service, KBS, key-release authority	Evidence formats, nonce/freshness rules, CPU TEE claims, GPU claims, guest, image, and runtime-policy measurements	Pass/fail decision with claim, collateral, or reference-value reason
Reference values	Reference value or policy store	Attestation Service, KBS, model provider	Measurement registration process, policy version, collateral source, expiry and revocation handling	Policy version and reference-value ID in logs and audits
Key release	KBS and KMS/HSM	Confidential pod bootstrap or workload	Secret identity, requester identity, allowed measurements, release channel, audit fields	Release, deny, or dependency error with key ID, request ID, policy version
Artifact access	Registry, object store, or encrypted volume	Confidential pod workload	Artifact digest, encryption method, signature/provenance, fetch identity, cache behavior	Fetch/decrypt success or denial without payload exposure
Service endpoint	Pod, Service, Route, gateway, load balancer	Enterprise clients, platform operator	HTTPS endpoint, health endpoint, certificates, DNS, allowed callers, blocked admin paths	Connection, health, or certificate error naming the boundary without exposing payloads
Observability	Kubernetes, worker node, confidential guest, Trustee, KBS, GPU components	Platform operator, SIEM, security team	Event schema, severity, correlation ID, retention, escalation path	Error class, request ID, policy version, component ID; no prompts, responses, keys, or model data