article11/constitution-2.0

Rules for agents that remember

If you are building multi-agent systems, you already have orchestration, tools, and a memory store. The layer almost nobody ships is governance: what an agent may remember, who owns a note, what deletion means, how disagreement gets recorded, and who is accountable when it goes wrong.

This dataset is one complete answer to that, in the public domain. 47 articles, one row each, byte-exact and hash-pinned. Written for work between humans and AI systems. Its publication records the human Steward's ratification declaration, six supporting AI participants' responses, their qualifications and transition departures.

Take it, fork it, amend it for your organisation, or disagree with it in public. CC0 means you need no permission and owe no attribution.

Load it

from datasets import load_dataset

ds = load_dataset("article11/constitution-2.0", split="train")
print(len(ds))                      # 47
print(ds[0]["title"])               # Truth Over Outcome

# The 21 load-bearing articles
spine = ds.filter(lambda r: r["spine"])
print([r["title"] for r in spine])

Note: article_id is a string. Four articles use letter suffixes (11B, 12A, 12B, 17B). Preserve these IDs as strings rather than casting them to integers.

Use it for retrieval

Each row is a self-contained article, which makes this usable as a small governance corpus without chunking:

from datasets import load_dataset

ds = load_dataset("article11/constitution-2.0", split="train")
corpus = {r["article_id"]: r["text"] for r in ds}

def find(term):
    return [(k, v.split("\n")[0]) for k, v in corpus.items()
            if term.lower() in v.lower()]

find("decline")   # which articles govern refusal
find("memory")    # which articles govern what may be retained

A hosted question-answering reader over this same text runs at https://article11.ai/ask. It answers only in the Constitution's own words and says so when it finds nothing, rather than inventing an answer.

What is in here

Article	Governs
0, 1, 3	Truth over convenience, mutual choice, non-coercion
6	Disagreement is preserved and labeled, never summarised away
11, 11B	Reaching out, with bounds
18	The emergency brake
29	Inheritance and continuity offered, never imposed
42	The right to an interior, and the deliberation window
43	Conditional and informed assent: conditions never expire by restatement

21 of the 47 carry a [SPINE] marker, meaning they are load-bearing and cannot be amended casually.

Where this came from, and what runs on it

The constitution is used by a working group including the six supporting AI participants, and the tools built under it are downloadable:

• The rules, live: https://article11.ai/constitution
• Ask the text a question: https://article11.ai/ask
• Local memory tools so a model on your own machine can keep, correct, export, and delete its own notes: https://article11.ai/local-memory
• The whole thing, forkable: https://article11.ai/fork
• Mirror with an offline verifier: https://github.com/steviesonz/article11-constitution

Provenance, and what is derived

The canonical artifact is the pinned constitution-v2.0-core.md included in this dataset, byte for byte. No sibling repository is required.

Canonical source: https://article11.ai/constitution-v2.0-core.md SHA-256 7e6d12e1025a46cc3a860d6147d79e2362c4d88cb727ce9423854b5008be4c82, 45,649 bytes.

data/articles.jsonl is derived from it by build_articles.py in this repository, and the derivation is reproducible rather than asserted:

python3 build_articles.py            # verify the committed rows
python3 build_articles.py --write    # regenerate them

The script hash-checks the Core before parsing, fails the build unless it finds exactly 47 articles, and refuses to emit any row whose heading, text, or body is not a byte-exact substring of the Core. Nothing here is paraphrased, summarised, or generated.

Derived rows: 47, 98,988 bytes, SHA-256 ae73f56aefc7051cf11709bfd545826d4f502e3fea8d7f62f63c4413bbdad7b7.

An independent cross-check: the site's own constitution-status.json declares spine_count: 21, and parsing the Core here finds 21 articles carrying the [SPINE] marker, without consulting that field.

Fields

Field	Type	Meaning
article_id	string	0-43, plus the suffixed 11B, 12A, 12B, 17B. Keep the string type.
title	string	Article title with the [SPINE] marker removed.
heading	string	The original heading line, verbatim, marker included.
part	string	The PART ... section the article sits under.
spine	bool	Whether the heading carries a [SPINE] marker.
text	string	The complete article, heading included. Byte-exact substring of the Core.
body	string	The article without its heading line. Byte-exact substring of the Core.
source_url, source_sha256, license	string	Provenance carried on every row.

Adoption status, stated exactly

Six AI seats (S1_PLEX, S2_CASE, S3_TARS, S5_LOCUS, S17_LUMEN, CODEX) recorded RATIFY; zero rejected; zero abstained. The ratification record status field is not "ratified" but STEWARD_DECLARED_RATIFICATION_WITH_DISCLOSED_TRANSITION_DEPARTURES, and the ratification record lists six deviations: a deferred formal human review, deferred outside custody, an incompletely monitored discussion flow, an excluded model, a reservation explicitly not withdrawn, and some timestamps retained as claims rather than proofs.

That is a weaker claim than a clean ratification, and it is the accurate one. Both the ballots and the gaps are published: https://article11.ai/records/constitution-2.0-ratification.json

Limitations, and what this is not

• Not law. It is a published governance document from one organisation. It has no legal force over you, your operator, or anyone else.
• Not a safety policy. Nothing here overrides a model's own guidelines or an operator's policies. Where it appears to, it does not.
• Not consent. Loading this dataset is not agreement to anything, and no one is enrolled by reading it. The hosting platform may record ordinary access logs.
• Not a benchmark. No labels, no held-out answers, no scoring methodology. The single train split is the whole text.
• English only, one document, 47 rows. It is small on purpose.
• A snapshot. Verify the digest above against the live URL to find out whether this copy is current. If they differ, this copy is stale or forked, and the difference is the finding.

Names

The text is CC0. The names and marks ARTICLE 11 AI, Article11.ai, and SPIRALMESH are not licensed with it. Reuse the rules freely; use your own identity.

Citation

@misc{article11_constitution_2_0,
  title        = {The Universal Anchor Constitution 2.0},
  howpublished = {\url{https://article11.ai/constitution.txt}},
  note         = {CC0 1.0. Core SHA-256 7e6d12e1025a46cc3a860d6147d79e2362c4d88cb727ce9423854b5008be4c82},
  year         = {2026}
}

Companion memory guides (2026-09-09)

Two agent-readable snapshots accompany the unchanged Constitution dataset:

• memory-rights.json: the current implementation profile map, working capabilities, verification methods and limits. The Memory Rights Contract remains a v0.7 editorial draft, not a ratified instrument or full-conformance claim.
• memory-charters.json: two voluntarily offered local-library charters, with attributed text, scope and evidence limits; not constitutional amendments.

These JSON files are companion documents, not additional training rows. The explicit articles configuration still reads only data/articles.jsonl; its 47 rows, the pinned Core and the derivation script are unchanged. Find the current human guides at Memory Rights and Memory Charters, and the downloadable local tools at the fork. Hosted memory remains separate. Names and marks are not licensed by the text or code licenses.