Skip to content

Compute and independent review options

Prepared 2026-10-20. Public documentation and pricing observations, not capacity reservations or evidence of working attestation. The metadata record preserves retrieved bytes' hashes and the public HF hardware inventory. Direct HTTP retrieval of the reviewer pages failed; the browser search reader supplied those pages. No authenticated API call, job, upload or reviewer contact occurred.

Ordinary research compute

Start on the owner's Mac: observed arm64, 48 GiB physical memory, 16 logical CPUs and about 610 GiB free disk. Use the approved four-thread/16-GiB/40-GiB envelope, two hours per job, six hours total. This is the selected execution route; it supplies no confidential-computing attestation.

Fallback Published hourly compute price Fit
HF T4 small, 16 GB VRAM $0.40 Lowest-cost CUDA smoke or small training job; 15 GB host RAM may constrain import.
HF L4, 24 GB VRAM $0.80 Preferred fallback if a measured local bottleneck requires CUDA or more accelerator memory.

Prices come from HF Jobs pricing; the API bills by the minute, with rounded minute rates. Starting and running time are billable. Jobs need a positive credit balance. The owner approved both T4 and L4 fallback options on 2026-10-21. Use at most one T4-small or L4 job, a one-hour provider timeout and a $2 total execution ceiling; no automatic retry, ports, scheduler, top-up or larger instance. This is an agent-imposed limit under the owner's HF Jobs authorization. Refresh the price and include starting time in the preflight estimate. Abort if the ceiling cannot be enforced.

Before launch, record the local bottleneck, source/data hashes, reviewed pinned container and locked environment, output location and cancellation procedure. Send only reviewed public inputs and source, with no private transcript, checkpoint or account secret in the command/log. Use the token for control-plane authentication; do not inject it into the worker unless a specifically reviewed input requires it. An ordinary HF GPU job supplies no evidence of a TEE.

Genuine confidential-computing hosts

Option Published capability Decision and missing evidence
Azure NCCadsH100v5 AMD SEV-SNP CPU with NVIDIA H100 confidential GPU. First candidate for a fixed, short qualification session. Region, quota, capacity, supported image and all-in quote still need account-specific checks.
Google Cloud A3 High a3-highgpu-1g Intel TDX with H100; Spot or Flex-start provisioning. Alternative hardware stack; interruption/queue behavior complicates reproducibility. Region/quota and all-in price are unconfirmed.
Google Cloud G4 g4-standard-48 AMD SEV with RTX PRO 6000; standard, Spot or Flex-start. On-demand alternative; includes an additional confidential-computing software licence charge. Review exact CPU threat coverage before choosing it.

These capabilities are described by Azure and Google Cloud. The Azure preference is an engineering recommendation, not a benchmark result. No cloud account, availability or price is assumed. Obtain an hourly quote plus storage/network charges for a two-hour qualification window before requesting a specific purchase. The owner asked for options, so no host is provisioned.

The qualification packet must include a fresh verifier challenge, validated CPU and GPU evidence and certificate/revocation state, pinned guest/model/runtime identities, a binding between that workload and both evidence paths, explicit debug-mode rejection, and tamper/expiry/replay negatives. Two successful isolated quotes do not demonstrate a bound end-to-end execution. Run this packet on harmless inputs before admitting any confidential workload.

Independent reviewer

Option Suitable scope Cost/availability
Owner-nominated independent researcher or engineer Reproduce the pinned run on a fresh machine, audit labels/splits and challenge the verifier against the negative corpus. Unknown until a person is nominated and agrees. Independence and relevant expertise must be recorded.
Trail of Bits professional assessment Candidate for a scoped security design and implementation review spanning trust enrollment, confidential execution and proof integration. Quote required; no fee or availability is assumed.

Trail of Bits' application-security service describes design and implementation assessment; its public reports provide examples of prior work. These references establish a candidate to assess, not an endorsement of this repository or confirmation of suitable staffing.

Prepare one review packet after the relevant implementation exists: frozen commit and dependencies, claims/limits, trust enrollment and threat model, data licence/lineage and label-audit records, reproducible commands, raw results, failure corpus and unresolved risks. Ask for an independently produced result with exact inputs, findings and remaining limitations. The authoring agent's self-check is not the independent review.

Recommendation: use the Mac now, HF only for a demonstrated bottleneck, qualify Azure's CPU/GPU evidence path when the local verifier is ready, and seek an independent reproduction before commissioning a broader paid assessment. Reviewer nomination and a concrete confidential-host budget are later owner choices; neither blocks current local implementation.