Who inherits a recipe?

Recipes have always moved between hands. A dish crosses a border in a bride's luggage, changes a spice because the market lacks one, and comes back decades later unrecognisable and loved. Nobody owns molokhia. And yet every version has a cook, and every cook has a name that the eaters know.

Cooking machines change the terms. A machine does not learn a dish from an aunt; it reads a file. Whoever writes the file decides what the dish is. If the files are written by a few companies in a few countries, then in a generation the machines of the world will cook a few cuisines well and the rest as approximations, and the aunt's version will be a video nobody can follow because nobody cooks any more.

Cookwala takes four positions.

A recipe carries the cook's name wherever it is cooked. The document has a source, a license and a signature. A machine in another country that cooks it inherits the attribution with the method. This is weaker than ownership and stronger than nothing: it is credit that travels.

A dish has an identity that a machine must respect. The identity section says what is essential, what is flexible and what is never added. A planner substituting for a budget or an allergy may change the flexible parts; if it changes an essential part, it is no longer making the dish and must say so. Koshari without crisp onions is a bowl of carbohydrates. This encodes a cultural judgment, and the judgment is the cook's.

The language of the cook is first class. Step text in Arabic is not a translation of the English; both are the recipe. A machine learning from execution logs learns the task as the cook named it.

Learning from a kitchen needs the kitchen's consent and credit. Execution logs leave a home only with opt-in consent, carry no personal data, and are published with a data card that credits contributing cooks by the name or pseudonym each chose, or collectively where they chose not to be named. A dataset that teaches robots molokhia should credit the cooks whose molokhia it was, on their terms; a credit is never a household's address, schedule or absence.

Three objections we take seriously. First, credit is not compensation; if a dataset of home cooking becomes a commercial model, attribution alone is thin. The standard's answer today is consent and a published data card; fair payment is a governance question for the foundation, not a schema field. Second, formalising identity freezes a living thing. Our answer is that identity in a recipe is a claim by that cook, versioned and forkable, not a canon. Third, the machines will cook what the files say regardless of what we write here. Our answer is that this is exactly why the files should be open and the format neutral.

The historian's question is the one we cannot answer: whether a cuisine survives better as working knowledge in machines or as practice in hands. We suspect both, and that the first without the second is a museum. Cookwala's part is to make sure that if machines cook the world's food, they cook all of it, and say who taught them.