Why Kiln

Java services already move tensors and fixed records between processes. A schema that is the struct itself keeps that path short: the field order is the wire order, and there is nothing to look up by name while encoding.

The struct is the schema

A generated class is the schema. encodedSize, encode, and decode walk fields in declaration order. Optional properties are a single presence byte. Dict keys are ordered by their UTF-8 bytes, so two encodes of the same map match. The dynamic API uses the same rules when the value is a Map or a long[] of fixed-width fields.

gld-kilngen-java --schema point.json --out src/main/java emits those classes from a JSON Schema document. The subset does not describe a tensor directory, so the generator does not invent a tensors field. A tensor block is an explicit directory value on the dynamic API.

Bytes stay bytes

A tensor directory describes dtype, shape, and quantization, then points at a payload. Decode does not widen f32 or f16 into a float[]. The payload is a slice of the input buffer. Numeric bytes are copied with one array copy on encode, not once per element. Negative zero and a quiet NaN payload stay the bits that arrived.

Hostile counts

A list count is a u32 that a peer can set to anything. Before the decoder allocates element storage it requires the count to be within the cap and count * minSize(element) to fit in the unread bytes. The same rule covers dicts and the tensor count in a directory header. The failure is LIMIT or EOF, and the element array does not exist yet.

The default cap is 16,777,216. A caller can raise it. Raising it does not skip the fit check.

Wire

The normative wire, including scalar layout, the frame hash, and the tensor directory, is specified only at:

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

This library is a Java implementation of that document. It is not the specification, and it does not restate the wire tables.