Skip to content

RFC-0047 — A domain article is sealed by the genesis commit

  • Status: Implemented
  • Track: I34
  • Relates to:
    • RFC-0024 (the policy layer, whose authoritative override is what a domain article must not be subject to, and whose "Composition" section named tighten-only and deferred where it is declared)
    • RFC-0026 (§3 item 2: a constitutional family's composition rule is declared in the Rust that owns the family, not in the policy)
    • #561 (2fcbd19, Article V's text: no exception to it is licensed "by a domain extension")
  • Versioning layers touched: template (bootstrap skill, CONSTITUTION.md's Domain extensions section, directories.md). Tooling: four yidam lint checks and a samudaya-audit warning. No SDK, parity or MCP contract change. No migration: a repository with no .yidam/constitution/ sees nothing new.
  • Downstream reference case: none yet. No derived repository has taken a constitutional augmentation. The two seed sets that carry one are samudaya/examples/genealogy/ and samudaya/examples/language-documentation/.

Summary

The constitution lets a derived repository add domain articles at genesis, permanently. Until now such an article was prose appended to a vendored file, and nothing checked it. This RFC puts each article in .yidam/constitution/, with an optional Rego rule and its cases beside it. yidam lint reads the rule from the genesis commit, not from the working tree. A repository that violates the rule fails lint. An edit to the rule after genesis is reported and not obeyed.

Problem

Three defects, found in order.

Nothing read the article. The Domain extensions section of CONSTITUTION.md licensed articles that "must be consistent with Articles I–VI". None of the Article V checks (#272–#275) read that section. The epic's test was "a resolution that violates the constitution fails CI". It held for yidam's articles and not for a derived repository's own.

The article did not survive a re-vendor. The bootstrap skill said to "append it to the repo's copy of the constitution". That copy is .yidam/.vendor/prelude/CONSTITUTION.md. yidam-vendor-update deletes that directory first (mise.yidam.toml: "rm -rf .yidam/.vendor/prelude"). So the first upgrade erased every article the genesis commit had appended. "Permanent" meant "until the next upgrade".

The obvious home is editable. A rule in .yidam/policy/ is authoritative under RFC-0024: a local file supersedes the rule for its package outright. An article there could be loosened by the corpus it binds, by one commit.

Proposal

1 — Where an article lives

Bootstrap copies a constitutional: true augmentation to .yidam/constitution/<stem>.md in the genesis commit. When the seed has <stem>.rego and <stem>_test.rego beside it, those are copied too. Nothing is appended to the vendored constitution.

The rule is ordinary Rego with one required shape: a partial set deny of messages. Its input is the sangha report (yidam sangha --format json) and the corpus's nodes as {file, class}.

2 — The genesis commit is the pin

yidam lint finds the root commit and reads .yidam/constitution/ from it with git ls-tree and git cat-file. The working tree is compared by blob id. A file whose id differs, or which is gone, is reported as domain-article-edited. A file added after genesis is reported the same way and never evaluated.

A hash file was the alternative. A hash file would sit in the tree, and the corpus could edit it alongside the rule. The genesis commit cannot be edited without rewriting history. A rewrite changes every commit id, so everyone holding the repository would see it.

A deleted article still binds. Its rule is read from the genesis commit whether or not the file is on disk.

3 — Composition is in the Rust (RFC-0026 §3 item 2)

Each rule is loaded into its own engine (Policies::sealed). The engine holds that rule and its cases, and nothing else. It loads no built-in defaults and nothing from .yidam/policy/. So no local file can supersede an article, because no local file is in the engine that evaluates one.

That is the composition rule, and it lives in the Rust that owns the family. RFC-0024 deferred the question of where; RFC-0026 answered it; this is the first family built under that answer.

4 — Tighten-only is the rule's shape (#561)

Article V says no exception to it is licensed "by a domain extension". The mechanism keeps that by construction. A rule answers only with refusals. A refusal fails lint. There is no allow a rule can set, so an article can add a refusal and cannot remove one yielded elsewhere.

A rule that cannot answer is not a permission. A deny that is undefined, or that errors, is reported as a violation at Error.

5 — Four checks

CheckSeverityFires on
domain-article-violatedErrora sealed rule's refusal, or a rule that could not answer
domain-article-editedErrora file in .yidam/constitution/ that differs from, or is missing from, the genesis commit
domain-article-unprovenWarna sealed rule with no sealed _test.rego, or a failing case
domain-article-unverifiableWarnarticles on disk with no genesis commit visible, such as a shallow clone

A shallow clone evaluates nothing. Its root commit is a graft, not the genesis commit, so reading rules from it would obey whatever the clone happened to start at.

yidam samudaya-audit warns before genesis when a constitutional augmentation has no rule beside it, or a rule with no cases.

What this does not do

  • It does not give a rule the whole corpus. The input is the sangha report and each node's file and class. The language-documentation article asks that a source state what it forbids, and a rule for it needs catalog frontmatter in the input. It ships as prose, and the audit says so.
  • It does not check that an article is consistent with Articles I–VI. The shape keeps an article from removing a refusal. Whether its prose contradicts an article is still the review samudaya-audit asks a person for.
  • It does not migrate an appended article. None exists. On 2026-09-29 all thirteen derived repositories on disk carried the template's Domain extensions section unchanged.