Security and Data

The core promise behind every Manar deployment is simple to state and meant to be checked, not taken on faith. Client data does not leave the building it was recorded in. The hardware running a model belongs to the client, not to Manar. The trained model itself belongs to the client once it is installed.

Offline is not a setting that gets switched on. It is a set of controls checked at two layers, physical and software, so that a mistake at one layer does not break the promise on its own.

Physically, unused network ports are disabled or removed, wireless radios are removed from the hardware where possible and blocked in software where they are not, and the machine is sealed with tamper evident marks logged at handover.

In software, the operating system is configured to block all outbound traffic by default, the model runtime is set to fail rather than silently reach outward, and every model file is checksummed against a manifest your own staff can recheck at any time.

At handover, your own staff run the check, not ours. A network scan confirms nothing is listening that should not be. The connection is physically removed while the system is actively answering a question, and it keeps answering. The firewall rules are reviewed on the machine itself, not described in a document. A written record is signed and kept by you, and the same check can be repeated at any point afterward, by your own people, without anyone from Manar in the room.

Once a deployment is running, Manar has no standing access to what it holds. Access only happens when you provide it, for a specific piece of work, at a specific time, with your own staff present. There is no telemetry and no remote session left open in the background.

The one honest limit is training. Building a model requires working with your data directly, because a model cannot be trained on data nobody has examined. That work happens on your terms, on your premises where you require it. What does not happen is a copy leaving with us afterward.