The model earns its tasks

There is an obvious way to ruin a sovereign-model project: move tasks to your own model because it is yours, before it is ready, and call the downgrade independence. A cheaper, worse answer in a triage is not a sovereignty win. It is a patient harmed in the name of a principle, which is the worst way to harm one.

So the rule is the opposite, and it is strict. A task moves from a frontier model to KhaM-Med only once KhaM-Med has been shown to match the frontier on that task, measured on evaluation sets built from the real thing: real Bangladeshi films from the actual machines in the actual centers, real dialect transcripts from real intakes, real chamber notes a real doctor wrote and corrected. Not benchmarks borrowed from abroad. The country's own work, scored honestly.

This is slower than flipping a switch, and it is meant to be. It means some tasks stay on a foreign model for a long time, because that is where they perform best for now, and the patient comes before the principle every time the two are in tension. Sovereignty is the destination. Safety is the constraint that sets the pace.

It also means that when a task does cross over, the claim attached to it is real: KhaM-Med reads this film as well as the frontier did, on films like the ones it will actually see. That is a sentence the network can stand behind, because it was earned on the country's own evidence, not asserted because the model wears the right flag. A home-built model that has to earn each task is worth more than one that was simply trusted, and it is the only kind that deserves to be.