The mapping model

Most data-transformation tools walk the input and decide what to emit. Mapper inverts that. A mapping declares the output shape, and evaluation walks the declaration: for each pairing, resolve the descriptor, write the result at the target pointer. The input is only ever read where a descriptor points. Nothing scans it. Unmapped input cannot leak, and cost follows the size of the mapping and what it addresses, not the size of the input document.

Two rules complete the model.

Undefined never writes. A descriptor that resolves to nothing leaves the output untouched. That is why array-descriptor fallbacks, find misses, and absent optionals compose without conditionals: the failed alternative simply writes nothing.

Order is the program. Pairings run in document order, and later pairings read what earlier ones wrote:

/a: { constant: 1 }
/b: { output: /a }

produces { "a": 1, "b": 1 }. Reverse the pairings and /b writes nothing: at that point /a does not exist yet. A mapping is closer to a small program than to a set of rules, and the pairing sequence is its statement order.

Structure descriptors descend. A nested mapping evaluates its pairings against a fresh target and writes the finished result back; each does that once per array element. Descent is what keeps scopes clean: inside the nest, pointers read from the narrowed source, and the parent output is reachable only by name (input:, output:).

The consequence to internalize: you read a mapping the way you read the output it produces, top to bottom. What a value is made from is written next to where it lands.