Evidence incomplete: cpu: Network error: prem-cpu-evidence HTTP 502 Bad Gateway; response artifact retained, not proof API is absent; gpu: Network error: prem-nvidia-eat HTTP 502 Bad Gateway; response artifact retained, not proof API is absent; Prem cpu_hardware_tcb_and_published_policy: Shared manufacturer verification plus fresh full report-data echo and published TDX invariants; failures/outages are retained separately. (unproven); Prem returned_gpu_evidence: Every returned detached GPU requires NVIDIA signature, issuer/time/nonce, set/digest and full NRAS security claims. Missing evidence/502 never passes. (unproven); Prem workload_measurements: Observed MRTD/MRSEAM/RTMRs are compared with the published illustrative baseline; it is explicitly unpinned, not a release/model-approved reference. Mismatches are recorded, not silently whitelisted. (unproven); Prem tdx_key_or_tls_commitment: This TDX endpoint echoes all 64 verifier-challenge bytes in report_data; it exposes no separate attested key/certificate commitment. Cloudflare TLS is not an enclave key. (unproven); Prem model_and_inference_route: x-session-id header observed=true; header presence and a model query are not signed routing proof. No inference was sent and no same-inference-connection binding was exposed. (unproven); Prem cpu_to_gpu_binding: CPU quote and GPU EAT are independent requests with independent nonces, with no joint cryptographic statement. (unproven); Prem exact_serving_gpu_set: EAT set equality covers every declared/returned GPU, not the complete physical serving inventory or scheduler/request assignment. (unproven)