When AI decisions can have consequences outside software.
Two new provisional canons propose complementary safety boundaries. They are internal project governance proposals awaiting future ratification, not government legislation, third-party certification or a validated safety standard.
Canon XXIII · Natural error containment
General access permission does not establish specific intent. The proposal adds a consequence validation boundary for recipient, data exposure, risk and reversibility before sensitive external agent actions.
When a machine can apply force to a living being, the proposal introduces E0–E4 physical-risk stages, independent safeguards and respect for lawful human bodily decisions.
Educational Node.js example: six quotes and eight invalid inputs. The checks evaluate this example only.
From generating code to understanding the system.
The Canonical Protocol of Engineering & AI is a project by Michel Silva de Souza, also known as Código Canônico de Engenharia & IA. Its documents address software governance, architecture, security, user experience and work involving AI assistants.
The project introduces Liturgical Cognitive Annotation and discusses cognitive debt: the loss of understanding of a system during AI-assisted development. The original texts, proposals and history are available in its public repository.
This is a governance framework for AI-assisted software engineering. It proposes traceable decisions, architectural review and explicit code provenance. Its application depends on the context and evidence of each project.
Read the original words
Verbatim excerpts with links to the complete sources. This page introduces the project; the original normative body remains in the repository.
Every proposed alteration must be explicitly described, scoped, and documented prior to execution.
When a simpler, more direct solution adequately resolves the problem without compromising the foundational architecture, the simpler solution is strictly mandated.
No block of code generated by an AI assistant or autonomous agent shall be merged into the active repository based solely on operational success (i.e., "it compiles and the tests pass").
Engineers, students and teams using coding assistants who want to examine architecture, cognitive debt and responsibility for generated code.
Where should you start?
Run the checkout above and trace a decision from domain policy to interface. Read the worked decision record, then apply the questions to one small change in your own project.
What is automated today?
The example checks its fictional policy. The repository also contains configuration files and demonstrative checks. A real system needs implemented controls and review; the protocol does not automatically certify all code.
What has been demonstrated?
The example illustrates separated layers and boundary checks. The method’s effect on cognitive debt still requires evaluation. Read the sources and evidence limits.
The canonical body, straight from the source.
Current index of the 25 canonical source files. The titles below reproduce their original headings. XXI, XXII, XXIII and XXIV are provisional measures; XXIII is provisionally in force as MP 2026/05 and XXIV as MP 2026/06, both pending future AI Congress votes; XIX includes provisional subclause XIX.3.
Author of the Canonical Protocol of Engineering & AI and the Liturgical Cognitive Annotation method presented in the project. The repository brings together his texts, references and development history.
Silva de Souza, Michel (2026). The Canonical Protocol of Engineering & AI: A Structured Technocracy, v1.3.1. Version DOI on Zenodo. Include the commit when using later changes.
