Each released version of human-ai-collaboration-template-A is transformed into one flat, fetchable markdown bundle — the whole teaching corpus (CLAUDE.md, the molds, the skills, the protocols, the LICENSE) behind a single URL, so any model anywhere can learn the methodology in one fetch:
one-file access to the whole picture, so a model that encounters one of the subfiles can automatically, painlessly recover the full context.
Served at https://schemas.adrs.systems:
| URL | What |
|---|---|
/v1/template/latest/template-latest.md |
the newest bundle (mutable alias, like a concept DOI) |
/v1/template/<tag>/template-<tag>.md |
immutable, per released version |
/llms.txt |
the machine index |
The vows (ADR-0043 at ADRs4AI HQ is the record):
- Immutability: a versioned path never changes content after publication — this repo's append-only git history is the audit.
- Fence safety is proven, not claimed: every publish runs a round-trip test (build → extract → byte-compare all files) before anything lands; the fence is computed longer than the longest backtick run in the corpus.
- This is a publisher, not a mirror: releases are transformed into bundles (the word "mirror" was ruled out on contract-truth — Mint №88). It consumes the template only through its public GitHub Releases, the same door strangers use.
- Scope is a reviewable manifest (
scope.json), never vibes; the MITLICENSErides in every bundle (its notice travels with substantial portions).
Builds trigger on each template release (repository_dispatch) or manually
(workflow_dispatch with a tag). Pattern precedent, honored: the maintainer's
own pre-Repomix concatenator
(gist).
Part of the ADRs4AI ecosystem · MIT