Trust & Security

What we run, where your data goes, and what we can prove.

This page lists what exists today, stated plainly. We are a design-partner-stage company: where a control is on the roadmap rather than certified, we say so. For security questionnaires (SIG-Lite, CAIQ, DDQ) write to andre@portmem.com.

Last updated July 2026.

Architecture & data flow

Where the product runs. The hosted tier runs on Google Cloud (Cloud Run + Cloud SQL, us-central1). Containers run as a non-root user. Health, uptime, and error alerting are monitored.
What leaves the boundary. In the hosted tier, document text is sent to Anthropic (verification judge) and NVIDIA (embeddings) under their API data-protection terms. Evidence anchoring sends only cryptographic hashes — never document content — to the OpenTimestamps calendar network.
Tenant isolation. Every tenant's data is isolated with Postgres row-level security, enforced at the database role level (the application role cannot bypass it), in addition to application-layer scoping.
Access. Console access is authenticated (Firebase Auth); API access uses per-tenant keys; administrative surfaces are role-gated.

Evidence custody — who holds the record

You custody your evidence; we do not archive it. Review evidence exports as a signed packet: hash-chained, Ed25519-signed, OpenTimestamps-anchored, and verifiable offline with our published key — with no dependency on PortMem continuing to exist. PortMem buffers packets briefly for handoff; the customer archives them under their own retention program (for example, a 17a-4 books-and-records system you already operate).
Verification key. 506afc81aa39331a4c704a97bca26100a0d857a682a6fa2dd7274b0dae547c3e — anyone holding it can verify our records; nobody holding it can forge one.
What we deliberately do not claim. PortMem is not a WORM archive and is not pitched as one. The design goal is that your records outlive your vendor.

Deployment tiers

Cloud (hosted). Fastest start; document text transits Anthropic and NVIDIA as described above.
Single-tenant. A dedicated engine instance per customer for corpus isolation.
Air-gap / on-prem. Local judge model and local embeddings: nothing leaves the customer boundary. Built for MNPI and data-residency requirements.

Subprocessors (hosted tier)

Google Cloud — application hosting, database, storage.
Anthropic — verification judge (Claude).
NVIDIA — embedding inference (NIM).
Vercel — this marketing site only; no customer document data.

The air-gap tier removes Anthropic and NVIDIA from the data path entirely.

Certifications & roadmap — stated honestly

Today. No SOC 2 certification yet, and we will not imply otherwise. The controls above are real and demonstrable in a technical review; an external security review of the platform was conducted in July 2026 and all findings were remediated.
Roadmap. SOC 2 Type I is on our certification roadmap. Design partners get full architecture and control transparency now, including a live walkthrough of the isolation and evidence path.