Why Kiln

Kiln is a binary layout for records and tensors. In this C library the struct is the schema. Field names are not written, tags are not written, and the bytes of field n + 1 start where field n ends. A reader that knows the declaration order can walk the buffer without a side table.

That is a good fit for C. An int32_t is four little-endian bytes. A list of float is a count and then the array. A tensor directory does not widen those numbers into a different container: the numeric payload stays a slice of the input buffer, which the caller can pass to a numeric routine as the array.

A hostile count is rejected before allocation. The decoder reads the count, checks it against the cap, then checks count * min_size(element) against the unread bytes. Only a count that passes those steps may allocate. An empty remainder with a huge count returns EOF and does not reserve a slot per element.

gld-kilngen-c --schema FILE --out DIR emits C structs for a JSON Schema subset. Each struct has encoded_size, encode, and decode. Optional properties are a 0/1 tag. This library is not the specification. It is one implementation of the wire.

---

The wire

Frozen values and the version-2 tensor directory are specified here, and only here:

https://leo-gan.github.io/gld-kiln/spec/

Scalars are little-endian. Strings carry a byte length and UTF-8. Dict keys sort by their UTF-8 bytes. A frame starts with KLN0 and a schema hash. A hash that does not match the destination type is SCHEMA. The directory magic is KLNT and the version byte is 2. This page does not restate the byte tables. The linked document is normative.