Specify the binary encoding, so Protocol Buffers is not required

[strimo378]

Clause 12 said a model "is serialized as a Protocol Buffers message" and left the rest to an editorial note, because Protocol Buffers has no citable specification with a stable identifier. That was one of the six blocking gaps: a standard cannot rest its wire format on a document it may not cite.

The clause now states the octets itself. An implementer needs this document and nothing else.

12.2	the four primitive encodings and their wire types: variable-length integer, 32-bit, 64-bit, length-delimited. Signed integers as the varint of their 64-bit two's complement, so every negative value is ten octets
12.3	the field key as the varint of (tag × 8) + wire type; order and repetition; the packed and unpacked forms of a repeated field
12.4	skipping an unrecognized field by wire type alone — which is why wire types 3 and 4 are excluded
12.5	limits
12.6	the message definitions, which stay in Annex B
12.7	the relationship to Protocol Buffers

Protocol Buffers moves from the normative references, where it could not be cited, to the bibliography. 12.7 says plainly that the encoding above is that of Protocol Buffers restricted to the constructs Annex B uses, that an implementation may therefore be built on a Protocol Buffers library, and that where the two differ this document governs.

The restriction is what makes the clause short enough to state at all: no groups, no zig-zag or fixed-width integer types, seven scalar types in all.

Annex B gains a wire type column, so the annex and the clause fit together — the tag and the wire type are what the field key is made of.

None of this is asserted

I wrote the clause from a reading of the format, not from a specification this standard is allowed to cite. That is only worth something if what the clause says is what implementations actually write. So scripts/check-encoding.rb restates the rules as a small encoder and compares its output, byte for byte, against the Protocol Buffers runtime:

• every scalar type the schema uses — including a negative int32, the largest uint64, a float negative zero, a string beyond ASCII
• both field key lengths, tag 1 and tag 20. Nine ONNX fields have tags above 15; I would otherwise have been stating the two-octet key untested
• both forms of a repeated field, and both forms of one field in one message
• the literal octets of the clause's own worked examples — an edit that makes an example wrong fails the build
• a nested message, by bootstrap: the script builds a message descriptor by encoding a FileDescriptorProto with its own encoder and hands it to the reference library, which rejects it or produces the wrong fields if the encoding is wrong. That is also where the high tags and the packed field come from — the library's own bundled types have neither

encoding: the clause and the reference implementation agree

28 of 28 comparisons agree. make check-encoding, run by CI in both workflows, so the clause cannot drift from what implementations write.

Also

google-protobuf is declared in the Gemfile. It arrived transitively through the Metanorma toolchain, and an undeclared dependency has broken this build twice already in this repository.

What is left of the gap

Only the maximum message size, now a minor row rather than a major one: the reference implementation applies 2 GB to a single message, and this document neither fixes that nor leaves it explicitly to the implementation.

One of the six blocking rows in Annex C is gone.

Verification

All three parts compile with "Syntax Valid", clean at severity 3 only — two new informational rows, both Relaton failing to resolve the new Protocol Buffers bibliography entry, the same class as the existing ONNX and JDF entries. check-schema, check-operators, check-encoding and check-stylesheet all pass. The rendered clause was read back from the HTML.

PDF is CI's, as usual — the STIX font host returns 403 in the authoring environment.

🤖 Generated with Claude Code

https://claude.ai/code/session_01DeGJcqqgBdz7kUpH6EsYxD

---

Generated by Claude Code