[art] WG: New I-D: JSON Structure: Semantic and Reference-System Annotations

[Clemens Vasters <[e-mail]>]

For awareness, following up on earlier discussions on JSON Structure.


________________________________
Von: Clemens Vasters <[e-mail]>
Gesendet: Donnerstag, 06. August 2026 17:03
An: [e-mail] <[e-mail]>
Betreff: New I-D: JSON Structure: Semantic and Reference-System Annotations

I-D: https://datatracker.ietf.org/doc/draft-vasters-json-structure-sem-ann/
Blog: https://json-structure.org/2026/08/06/semantic-annotations.html

Hello,

this is the 7th companion layer to JSON Structure Core, and first spec that I've written where the primary audience for the defined metadata annotations are AI/LLMs. The goal of this spec is to make schemas semantically precise, in several dimensions, so that LLMs can write more correct code and queries related to the described data.

Two schemas can agree on "double", agree on the unit "m", pass every validator, and still be describing water level above a tide-gauge datum and height above the ellipsoid. Subtract one from the other and the answer is not a number about anything. The fact that would have stopped that subtraction lives in a PDF, or in a field name somebody hopes will be read the right way, because the schema language has nowhere to put it.

This new I-D gives the information somewhere to go. It extends JSON Structure with optional annotations recording what a value is read against: the vocabulary term, the observation role, what has already been done to the value, the temporal regime, the coordinate or linear reference system and which of your members supplies which axis, the vector and tensor frames, the color space, the channel layout, the spectral bands, the code register, and the weighting curve behind a decibel.

The I-D does not define any of those systems—other bodies do that, and an annotation points at one. What it defines is the form of the pointer and the rules by which a processor checks that your members agree with what you pointed at.

One test decided every keyword: a consumer must need it to decide whether two values may be combined, and it must hold for the type rather than varying per record.

There are 43 worked samples, 28 of them transcribed from feeds that actually exist. There is also an AI/LLM evaluation harness, and two recorded runs are in the repository along with the caveats that limit them.

What I would find most valuable now is, along with any other feedback, is whether data sets of yours have semantics, reference systems or structures that the keyword set cannnot express.

Draft: https://datatracker.ietf.org/doc/draft-vasters-json-structure-sem-ann/
Repo: https://github.com/json-structure/semantic-annotations
Harness: https://github.com/json-structure/semantic-annotations/tree/main/evaluation
Samples: https://github.com/json-structure/primer-and-samples/tree/main/samples/semantic-annotations
Write-up: https://json-structure.org/2026/08/06/semantic-annotations.html
Home: https://json-structure.org/




Clemens



[cid:91708a8a-aad0-4c0b-802d-f3b5f5814919]

Clemens Vasters

Messaging Platform Architect

Microsoft Azure

È+49 151 44063557

•  [e-mail]<mailto:[e-mail]>
European Microsoft Innovation Center GmbH | Gewürzmühlstrasse 11 | 80539 Munich| Germany
Geschäftsführer/General Managers: Keith Dolliver, Benjamin O. Orndorff
Amtsgericht Aachen, HRB 12066