Bounded by design
Signed job envelopes, immutable commitments, expiry, resource ceilings, and replay rejection establish the scope before a provider runs work.
02 / Technology
A control plane for verifiable compute.
The architecture treats execution, verification, recovery, and settlement as separate decisions. That separation helps a buyer inspect each control instead of accepting a single blended trust claim.
Policy selects the proof
Verification can be workload-specific: deterministic replay, redundant independent execution, attestation interfaces, or a future hybrid path.
Runtime Execution Evidence before value
Customer charge, provider payout, network fee, and any future issuance remain separate facts. Verification must succeed before settlement eligibility.
System model
From an approved request to a reviewable result.
The intended flow keeps arbitrary customer shell execution out of the trusted worker path and preserves the decisions that matter to later review. The Architecture describes system components; the Verification Journey describes the runtime decision flow.
01 / REQUESTDefineCustomer creates a bounded job request.
02 / PROVIDERScheduleEligible provider is selected under Runtime Execution Evidence and policy constraints.
03 / VERIFICATION DATAVerifyCompletion data and result commitments are evaluated.
04 / RECOVERYContinueBound checkpoints can support an authorized recovery path.
The technology direction is isolated from the deployed validator runtime while development and review continue.
Authoritative reading
Architecture is the map. Evidence is the decision.
Use the technical architecture for the full system boundary, and the assurance documents for current proof and open gates.
Recorded result, bounded claim
Attestation identity is not CUDA execution proof.
The recorded Google Confidential Space + Intel TDX + NVIDIA H100 run verified signature, GCE, GPU, and workload identities and observed a CUDA probe. The result is PASS_NON_PROMOTABLE, classified as ATTESTED_WORKLOAD_AND_GPU_CORESIDENCE, not authenticated proof of specific CUDA execution or output.
Strict execution/output binding is BLOCKED_BY_VENDOR_PRIMITIVE_AVAILABILITY. No supported primitive/API was identified in documented public NVIDIA and Google interfaces reviewed as of October 2026. Additional internal testing alone cannot establish Tier 3; a vendor-supported mechanism or newly documented supported path requires separate qualification. The gate remains closed and production authorization is not granted.
Agent Trust Runtime v1.0 is an additive agent-policy layer with 13 local/synthetic scenarios, not H100 hardware evidence or a production backend. The public continuity/provider-failover demo is a separate mock-fixture record.