Datasette's MCP server swapped rows for objects

Simon Willison changed how his datasette-mcp tool returns SQL query results on September 1, swapping arrays for objects. Rows used to come back as arrays of arrays, values in column order; now each row is an object with column names as keys. A model that loses track of which position maps to which column doesn't throw an error. It attributes the wrong value to the wrong field and returns a wrong answer that looks plausible.

Give a model a positional row like [42, "Portland", true] and it has to hold the schema in its head. It needs to track which value maps to id, city, and flag.

The trade-off is bandwidth. Repeating column names on every row costs more tokens than sending them once and letting position carry the meaning. Willison's choice says that cost is worth paying against a wrong answer that looks right.

Most MCP servers returning tabular data default to arrays of arrays, because that's what a database driver hands back. Willison's data point applies specifically to weaker models. A frontier model might handle positional rows fine, but a smaller one won't flag the mix-up at all.