Skip to content

Translation

Translated from the Spanish source, which is the normative text. Status: draft pending human review. Differences are resolved in the Spanish text.

AI that maintains this repository

This repository documents methods for working with AI and is also maintained with the help of AI assistants. This page explains how the same methods apply to that work: who may do what, with which permissions, what evidence they leave and which person keeps authority.

Status

The MANT-AX axioms are maintenance instructions for this repository. They are not ASILO-R, MILPA-R or SCAFFOLD-R requirements, they do not modify any accepted edition and they do not bind anyone who adopts the methods. Making them normative requires a decision recorded under the governance.

Three levels and where recursion stops

Level What it is What governs it
0 · Method The accepted text of ASILO, MILPA and SCAFFOLD. Decisions recorded under the governance.
1 · Project This repository as a documentation project. The methods themselves: SCAFFOLD for work, branches and releases; MILPA for sources and contradictions; ASILO for permissions, data and effects.
2 · Operation The people and assistants who read, draft, translate, review or integrate changes. The MANT-AX axioms and the controls inherited from level 1.

Evaluation stops at level 2. No agent is created to supervise the agents, and no method is created to review the review method. Assistant profiles are repository files: they are proposed, reviewed and integrated like any other level 1 change. Tools (MkDocs, Python, AI models, continuous integration) are treated as dependencies with a declared version, not as subjects of a new recursive evaluation.

Inherited controls

Each level 2 assistant inherits, and cannot widen:

  • The effective permissions of the host tool and session: files, network, execution and approvals. An agent profile describes responsibilities; it grants no permissions.
  • Change review in Git: branch, diff, PR and the maintainer's decision.
  • The repository checks (scripts/check-all-docs.py and tests in tests/).
  • The integrity records of accepted editions in reference/.
  • The authority of the person responsible for each decision, according to the governance.

Maintenance axioms

MANT-AX-001 — Declared identity and task. Whoever works declares their function (profile), the goal of the task, the sources and paths used, their limits and the responsible person who assigned the work. Related: SCAFFOLD-R-009 and SCAFFOLD-R-030.

MANT-AX-002 — Facts kept apart from everything else. A report distinguishes verified facts, inferences, translations, proposals and accepted decisions, and does not present one as another. Related: MILPA-R-002.

MANT-AX-003 — Least privilege and prior approval. Only the permissions the task requires are used. Before an external, destructive, legal or irreversible action (push, publication, deletion, licence or access change) human approval is obtained for that specific effect. Related: ASILO-R-004, ASILO-R-005 and SCAFFOLD-R-010.

MANT-AX-004 — Protected material stays out of the work. Secrets, personal data, confidential material and third-party data are not read, copied or published. Examples are fictional. Related: ASILO-R-006.

MANT-AX-005 — Reviewable evidence and real checks. Every change stays in Git with its reason and the checks that were run. No one reports a check that was not run or a human review that did not happen. Related: SCAFFOLD-R-003 and MILPA-R-001.

MANT-AX-006 — No one approves their own work. An assistant does not approve its own normative, licence or normative-term translation changes, nor compliance conclusions. Those changes require competent human review. Related: SCAFFOLD-R-022 and SCAFFOLD-R-023.

MANT-AX-007 — Bounded recursion. Applying these axioms ends at level 2. A question about an inherited control is escalated to the maintainer instead of creating another level of supervision.

AI use in external contributions

You may use AI assistants to prepare a contribution. State in the PR which part the AI produced and what you reviewed, as the template asks. Whoever submits the PR is responsible for its content. Do not attach private conversations with an assistant.

Repository profiles

The assistant profiles and their shared contract are at the repository root (AGENTS.md and each tool's configuration folders). They are not part of the published site, and their presence does not mean an assistant is running.