11.9 How the model reaches the endpoints
11.9.1 What is in the signed policy
The server adds two things to the signed policy it serves to agents (they are not in the editable policy and cannot be changed in the console):
| Item | Content |
|---|---|
| AI model reference | Model version, file names, sizes and SHA-256 hashes of the encoder and tokenizer. Present only when the model files on the server are complete. |
| Decision layers | Your trained layer (weights only — no example text), only if it was trained with the current model version. |
Both are removed from the public variant served to web add-ins.
11.9.2 Download and verification
- An agent whose scope has the AI module on downloads the two model
files once over port 443 from
https://ada.example.com/ai/…(about 57 MB). - A file whose size or SHA-256 differs from the signed policy is discarded and never loaded.
- Verified files are kept in the agent’s protected data folder (Windows: SYSTEM and Administrators only); older model versions are deleted once the new one is ready.
- Failed downloads are retried at growing intervals (1 minute up to 1 hour).
The agent’s status window shows the state as AI Classification: Ready, Preparing (downloading or verifying), Unavailable (last attempt failed) or Off (module off in the policy) — in the Windows tray About window and the macOS menu bar.
11.9.3 When endpoints switch to new results
| Change | Publish needed? | When endpoints use it | Side effect |
|---|---|---|---|
| Training completes | No — the server re-signs automatically | Next policy sync | Files are re-evaluated once (below). |
| Type deleted (last type) | No | Next policy sync | AI scoring stops. |
| Suggested label changed | No | — (affects only new rule drafts) | None. |
| AI rule created/edited | Yes — Publish & Sign | Next policy sync | Re-evaluation if the AI part of the rules changed. |
| AI module switched on | Yes | Next sync; then model download | First full re-evaluation. |
| Product upgrade with a new model | Server upgrade | After you retrain | Until you retrain, AI conditions do not match (section 11.9.4). |
Policy sync follows the scope’s sync interval (default 300 seconds).
11.9.4 Model versions and upgrades
The AI model changes only with a product release that ships a new model. After such an upgrade:
- Your document types shows The AI model was upgraded since the last training — train again so agents receive your types. Until then your layer is not offered to agents and AI conditions do not match.
- Click Train. Stored examples are re-read with the new model from their encrypted text; no re-upload is needed.
Upgrade the server before the agents, as for every release.