Socair (SUH-ker), Irish for at ease, settled, secure
Know what you are about to serve.
Model assurance for teams that run open-weight AI on their own hardware, on-prem or air-gapped.
Socair inspects a model artifact and produces a signed report of what was checked in those exact bytes, what was found, and what was not tested. Anyone can verify it offline.

Why Socair
-
The weights come from strangers.
You pull a model built by people you have never met, and serve it on hardware you own, next to data that matters.
-
Checking them should leave a record.
Not a green dashboard. A signed record of exactly that file, which your security team can keep and anyone can verify offline.
-
And the record should say what it did not check.
Every honest claim has a limit. Socair states its own, every time.
How a model comes in
One intake host brings each model in through the airlock, a store with a staging area and a clean area. Your GPU servers mount only the clean store, read-only.
-
Step 1:Pull it at a pinned commit. Every file is checked against the hub’s own hash.
-
Step 2:Scan it, offline. Each Tier 1 check says what it found, or that it did not look.
-
Step 3:Sign the report. The issuer is whoever signs, with their own key.
-
Step 4:A named person accepts any untested gap, with an expiry. A FAIL or a LEAD cannot be accepted.
-
Step 5:Promote it through the airlock into the clean store. The log records it. Every step, with its commands
The console’s other pages, Pending and Activity, are with the airlock, step by step.
What a report claims, and what it does not
The claim
For the artifact identified by its hash, the Tier 1 checks that returned PASS found no indicators within their stated scope. Rows marked NOT_TESTED were not examined, for the reason each row gives. Tier 1 runs no inference, so no model behavior was tested. Every surface outside that scope is listed under Out of scope.
One of the two fixed sentences a Tier 1 report carries, chosen from its rows. A report with a FAIL or a LEAD can never carry this one.
Both fixed sentencesThe limit
This attestation does not certify the absence of unknown backdoors.
- Backdoors, trojans, or poisoning in the model weights. Tier 1 reads the weights' layout, never their behavior.
- Triggered, sleeper, or polymorphic behavior. Tier 1 runs no inference.
- Behavioral safety: jailbreak susceptibility, harmful capability, and bias.
- Artifact formats we do not parse.
Four of its lines. Every report carries that sentence and the whole list under Out of scope.
Read the full ceilingOpen source, every check
- The engine and every check, the report format, signing and verification, the airlock, and the console: free under Apache-2.0, with every check included.
- Release builds are reproducible. Every binary carries a build-provenance attestation signed through Sigstore and, from 0.1.0-rc.4, an attested SBOM.
- The engine’s dependencies are vendored, so it builds air-gapped.
Contact
Running open-weight models where the network can’t reach? Write to hello@socair.ai
Found a vulnerability? Report it privately through GitHub’s private vulnerability reporting, or write to security@socair.ai.