ardregistry.net

Reference

ARD specification changelog

The specification is a proposal and it moves. Most published coverage predates the August revision and is wrong about the manifest path.

The change that matters: v0.91, on 26 August 2026, renamed the manifest from /.well-known/ai-catalog.json to /.well-known/ard.json and replaced rel="ai-catalog" with rel="ard". The same revision restated entries as JSON-LD nodes, which adds a @context seam without invalidating anything already published.

Why this page exists

The specification is a proposal and it moves. The August revision renamed the manifest, which means most of the coverage published in June is now wrong on the single most important detail. Nobody is going back to correct those posts. This page tracks what changed so you can tell whether something you read still applies.

How this is compiled. Compiled by reading successive revisions of the published specification. It is not an official changelog and the working group does not maintain one at the time of writing. If something here is wrong, it is ours to fix.

v0.91 · 26 August 2026 · status Proposal

The manifest was renamed

BeforeAfter
Manifest path/.well-known/ai-catalog.json/.well-known/ard.json
Link relationrel="ai-catalog"rel="ard"
Consumer dutyn/aMUST fetch ard.json, MUST honour rel="ard". Consulting the predecessor is optional.

The practical consequence is stated plainly in section 5.1: a resource that remains only at the predecessor path may not be found, because consulting it is optional for consumers. A publisher on the old path should move. Ten-minute migration.

Entries became JSON-LD nodes

An entry is now a JSON-LD node whose plain terms resolve through a base context published at https://agenticresourcediscovery.org/context/v1, mapping to the default namespace https://agenticresourcediscovery.org/ns#. An entry may declare its own @context to draw terms from additional vocabularies, and those terms become available as filter dimensions with no further spec change.

Nothing published before this breaks. The specification says so directly, calling it a repositioning rather than a redefinition, and carrying @context in an entry is optional. You want it when the entry will be read by generic JSON-LD tooling that has not been told it is looking at ARD, which in practice means in-page markup.

Queries gained a context too

The query object in 5.3.1 now takes @context alongside text and filter, so a filter key naming a namespaced term resolves to an IRI rather than being matched as a literal string. That is what lets a client filter on okf:taxonomy and match an entry whose author bound the same namespace under a different prefix.

The catalog entry and the ARD entry were separated

Section 4 now states explicitly that the two are different objects and that their conformance is self-contained: a change to what a catalog entry requires does not change what an ARD entry requires. This is the formal version of the distinction that explains why a valid manifest can be unfindable.

Original release · 17 June 2026

Announced by Google, with launch coverage from Microsoft, AWS, Hugging Face and Cisco the same week. The manifest was ai-catalog.json. Built on the AI Catalog data model from the Linux Foundation's AI Catalog Working Group, licensed Apache 2.0.

What to watch

Last reviewed 2026-09-04. Checked against ARD v0.91 (Proposal, 2026-08-26).