Using an admitted model

Admitted means the gate found nothing it could not settle, for that exact revision. It does not mean the model is good for your work, and it does not mean the model is installed. Heliast does not host weights.

1. Get the weights from the publisher

Each verdict page names the repository and the revision the gate assessed, and the digest of the weights. Download that revision. A different revision is a different model as far as the gate is concerned, and it has no verdict until it is assessed.

2. Pull and verify

Your operator stages the weights with the pull tool, which checks every file against the digest the gate pinned:

model-pull pull \
  -manifest <your weights-manifest.json> \
  -source-url <the publisher repository, or a mirror you control> \
  -dest <your models directory>

The tool checks the files by digest, checks the format, and checks a signature when you pin the publisher's key. If a file does not match, the pull stops, nothing is placed, and the failure is recorded as a security event. Files are placed only when every file of every model in the set has passed.

The model needs an entry in your weights manifest. The manifest is written from the admitted records. If the model you want is not in yours, ask your Heliast contact.

3. Check it fits your hardware

The verdict page shows how much memory the weights need and which of the five hardware tiers can hold them, with room left for conversations. This is memory arithmetic, not a speed claim. See hardware for what each tier is, what has been measured on it, and why the room left for context decides how many conversations run at once.

4. Entitle it in your policy

A model nobody is entitled to cannot be used. Add the model to your policy config, with your own group:

{
  "model_digests": { "<model id>": "<weights digest>" },
  "model_attestation_versions": { "<model id>": "<key id>|<issued at>|<gate version>" },
  "entitlements": { "<your group>": { "models": ["<model id>"] } }
}

The console fills in the digest and the attestation version from the model's record, so the two cannot disagree. The enforcement hook refuses a name it cannot match to an admitted record. See policy and entitlement.

5. Check it on your own tasks

Run the model on your tasks in Compare. See Compare.

If the verdict changes

A verdict can move. If a model's verdict becomes held or rejected, it stops being admitted. Once your estate has the updated record, the hook stops letting it through, and the ledger shows the denials. See reading a verdict.

Limits

• The pull and policy steps are described from the tools as built. They have been tested, and have not yet been run end to end in a customer estate.
• A model you trained or tuned yourself is not part of public admission, and the console never shows it as admitted.