Sovereign by construction.
Security here is not a compliance page. It is the architecture: post-quantum sealing by default, keys that never leave your enclosure, lineage signed on every object, and zero required outbound calls. Named algorithms, cited standards, no adjectives.
Assume the computers of the 2040s
Data sealed today must survive adversaries who record now and decrypt later. Midgard therefore encapsulates keys post-quantum by default, no legacy mode, no opt-in checkbox.
| Layer | Algorithm | Standard |
|---|---|---|
| Key encapsulation | ML-KEM-768 (default) · ML-KEM-1024 (high assurance) | NIST FIPS 203 |
| Symmetric sealing | AES-256-GCM | NIST FIPS 197 · SP 800-38D |
| Signatures, lineage & releases | ML-DSA-65 | NIST FIPS 204 |
| Hashing, content addressing | SHA-256 / SHA-3-256 | NIST FIPS 180-4 · FIPS 202 |
An ML-KEM-768 public key is 1184 bytes. If yours isn't, the platform tells you the fact and the fix: regenerate with midgard keygen.
What never leaves
Generated inside the enclosure, held by you, rotated by you. Unsealing is an audited event. Veldris cannot read your objects, by construction, not by policy.
Mimir runs models locally, pinned by hash. No prompt, no embedding, no telemetry crosses the boundary. Air-gapped operation is a supported mode, not a degraded one.
Zero required outbound calls. The platform is fully functional with the cable unplugged; anything optional that talks out is off by default and named in the source.
Lineage is the security model
Every object carries a signed, append-only lineage: what produced it, from which inputs, under whose authority. Every access decision the platform makes is logged against that line.
This is why a breach investigation in Midgard is a query, not an archaeology project, and why a claim on this website can carry a provenance tag instead of an asterisk.
The supply chain holds the same standard. Builds are reproducible; releases ship a signed SBOM; Bifrost refuses to deploy anything that cannot prove its origin, on either side of an air gap.
And because the core is open source, none of the above asks for your trust. Read the source →
Found something?
Report vulnerabilities to security@veldris.io. State what you found and how to reproduce it; we respond with the fact and the fix, and we credit reporters who want credit. No legal threats for good-faith research, in writing.
AES-256-GCM under ML-KEM-wrapped keys. Read the source.