RFC 0021: Named Dialect artifacts

• Status: Accepted
• Accepted: 2026-10-16 (effective upon merge of the adopting pull request)
• Type: Standards track, draft maturity
• Created: 2026-10-16
• Issue: #180

Summary

Give a reusable input Dialect an independent id and three-component artifact version, and select its document grammar through formatVersion. Keep integer formats and receipts intact.

Problem

A recipe edit counter cannot identify a portable release. Replacing integer version 2 or 3 with an artifact version conflates grammar with content and changes digests without specifying readers. Jobs and catalogs need an immutable release identity while drafts and research records keep changing.

Evidence

Desk's company-qualification flow edits a translation while jobs retain the previously reviewed mapping. A healthcare procurement recipe needs the same lifecycle without company-specific fields. Runner already freezes bound documents, and the two repository preparation readers distinguish format 2 from 3. Runtime already provides exact shared version arithmetic. None establishes a portable Dialect artifact identity today.

Specification

Add §2.1 of the Dialect companion. Its exact selectors are 2.0.0 and 3.0.0. They retain both translation families and have new canonical identities. The preparation formats stay integer 2 and 3. Document validation, publication admission, installation binding and execution remain distinct checks.

Alternatives

• No change leaves reusable identity in each product's private metadata.
• An extension or profile cannot add members to the closed document grammar or change its digest.
• A release envelope could retain the old document unchanged, but makes the identity indirect and requires every reference to distinguish envelope and content without giving authoring a root identity.
• Reusing the integer version or accepting string aliases silently changes historical identities.
• Product-only semantic versions do not make independently exchanged documents self-identifying.

Compatibility

Old readers refuse new documents; writers discover exact supported selectors before emitting them. Old integer documents and all recorded vectors remain byte-identical. Conversion is an explicit new-artifact operation, never a receipt migration. New family-2 documents use the one-way encoding rule already used by family 3; legacy empty root fields do not survive that explicit conversion. Output and pack-free presets remain separate, proposed contracts; this amendment does not relabel them.

Security and privacy

An id is bounded opaque data, not a URL to fetch. Identity is not authenticity or permission. Catalogs reject identity reuse with different canonical content. Bound derivatives keep separate provenance. A valid document cannot authorize acquisition, infer credentials, bypass review, or silently upgrade an existing job. Strict JSON, integer canonicalization and byte bounds still apply.

Conformance

Add staged positive, malformed, unsupported and binding tests to both repository readers. Exercise both families, exact canonical digests, independent artifact versions, no coercion, leading-zero refusals, changed identity, bound derivatives and historical vectors. These are research-preview agreement tests, not a published preparation-verifier conformance claim.

Implementation

Python and TypeScript read the new headers and retain the original document for canonicalization. Runtime exposes format capabilities and authoritative checks; Runner keeps frozen named documents and installation bindings. Desk reuses shared review/publication/version controls. These changes are by one author with model assistance; two language implementations are not independent authorship.

Unresolved questions

An output-side release and a pack-free preset need separately adopted contracts before sharing an artifact identity. Research-source execution remains subject to the accepted source contract and reader capabilities. This amendment does not implement a hosted registry or storage adapter.