Secure Research AI Cluster

An on-prem GPU environment for AI on sensitive and regulated research data. As a local research deployment, it leans on a system card, internal model summaries, and risk evidence rather than a public service card.

Fictional example. The Secure Research AI Cluster and its configuration are invented to illustrate how the framework documents a high-sensitivity research environment.

See them as documents. This tool's internal artifacts are available as formatted, printable documents — the System Card, the Internal Model Summary, and a Confidential risk-evidence review — so you can see what restricted and confidential documents look like. Use the View as document links below.

Access summary

Catalog entry Restricted Internal

A short, shareable summary lives in the catalog so researchers know it exists and how to request access. The detailed documentation below is internal.

Audience
Approved researchers
Data allowed
Confidential & Regulated
Access via
Research Computing request
Owner
Research Computing

System card

Layer 5 · System Restricted Internal View as document 
Document Title:    Secure Research AI Cluster — System Card
Document Type:     System Card
Primary Audience:  Reviewers, security, technical owners
Visibility:        Restricted Internal
Owner:            Research Computing
Approver:         AI Governance Committee + ISO
Version:          1.2
Review Date:      2026-03-30
Architecture
On-prem GPU nodes in an isolated network segment. Open-weight models run locally — no data leaves the cluster boundary. Job scheduling and storage are managed by Research Computing.
Identity & access
Per-project allocation, multi-factor authentication, and least-privilege roles. Access is time-boxed to the approved study period.
Controls
Network isolation, encrypted storage at rest, full audit logging, data-egress prevention, and per-project data segregation.
Governance boundary
Approved for Confidential and Regulated data under an IRB protocol and data-use agreement. General-purpose chat use is out of scope — that belongs on Meridian GPT.

Internal model summaries

Layer 6 · Model Restricted Internal View as document

Because the cluster runs open-weight models locally, Meridian maintains internal summaries that capture intended use, limits, and local deployment context.

ModelApproved forDocumented limits
Open-weight LLM (large)Text analysis on sensitive corporaHosted locally; quality below frontier hosted models
Open-weight LLM (small)High-volume batch processingWeaker reasoning; validate on a sample first
Domain embedding modelSearch and clustering within a datasetEmbeddings are dataset-specific; not portable

Supporting evidence

Layer 7 · Evidence Confidential View as document

High-sensitivity environments carry the deepest evidence trail. These are confidential and listed only to show they exist.

  • Security architecture review and penetration test results
  • Per-project IRB protocols and data-use agreements
  • Data-flow diagrams for each approved study
  • Risk register with residual-risk ratings and mitigations
  • Access-recertification and audit-log review records