You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Run workloads inside hardware enclaves (Intel SGX, AMD SEV) so that a provider hosting a workload cannot read its memory or data. This is the structural answer to the Byzantine-provider problem that stake and reputation only mitigate: ADR-012 §2 lists operator-level collusion as unsolved, and confidential compute is what eventually closes it.
Blocked by: ADR-019 (see ADR-012 §6) — no implementation before acceptance. #61 provides the proof that this issue's guarantee is real; the two ship together or the guarantee is unverifiable.
Acceptance criteria:
ADR-019 states which vendor roots of trust are accepted, the revocation path when one is broken, and what an unattested provider may still do;
TEE capability is advertised as a distinct, verifiable resource property, and a tenant can require it as a scheduling constraint;
a workload requiring a TEE never runs outside one — failure to place is an explicit scheduling failure, never a silent downgrade to plain execution;
the trusted computing base is documented, including what the provider can still observe (scheduling, traffic volume, timing) despite the enclave;
known TEE limitations — side channels, rollback attacks, vendor key compromise — are documented as residual risk rather than claimed to be solved;
TEE support is opt-in per workload and never network-wide, preserving the rollback property required by ADR-012 §7;
tests cover placement on an attested provider, refusal to place when none is available, a provider falsely advertising TEE support, a revoked vendor root, and secret release denied on failed attestation.
Original request (FR, preserved verbatim):
Permettre l'exécution des workloads dans des enclaves sécurisées pour garantir la confidentialité des données.
Run workloads inside hardware enclaves (Intel SGX, AMD SEV) so that a provider hosting a workload cannot read its memory or data. This is the structural answer to the Byzantine-provider problem that stake and reputation only mitigate: ADR-012 §2 lists operator-level collusion as unsolved, and confidential compute is what eventually closes it.
Blocked by: ADR-019 (see ADR-012 §6) — no implementation before acceptance. #61 provides the proof that this issue's guarantee is real; the two ship together or the guarantee is unverifiable.
Acceptance criteria:
Original request (FR, preserved verbatim):