Translation
Translated from the Spanish source, which is the normative text. Status: draft pending human review. Differences are resolved in the Spanish text.
ASILO and MILPA principles — editions 0.1¶
Current editorial presentation of the twelve accepted principles. Statements and IDs remain unchanged; explanatory material preserves their scope.
Informative translation. The Spanish text is authoritative. If the two differ, the Spanish version prevails.
A principle guides decisions; it is not checked directly. The requirements translate each principle into verifiable obligations. All principles on this page are agnostic: they do not depend on language, platform or model provider.
MILPA¶
| ID | Principle | Derived requirements |
|---|---|---|
| MILPA-P-01 | Authority is declared and recorded; it is not inferred. Location, file name or a canon label do not confer authority on their own. |
MILPA-R-001, MILPA-R-003 |
| MILPA-P-02 | Facts, inferences and hypotheses are distinguished before a claim is reused or acted upon. | MILPA-R-002 |
| MILPA-P-03 | A small core and a declared profile. The method fixes what every project shares; the profile fixes what depends on its type. | MILPA-R-005, MILPA-R-006 |
| MILPA-P-04 | A project's identity can be read before exploring it. Who is responsible, what it covers and what is restricted should not have to be deduced by walking through folders. | MILPA-R-004 |
| MILPA-P-05 | Relationships between projects are declared at both ends. A similar name or a shared folder is not an integration. | MILPA-R-007 |
| MILPA-P-06 | A contradiction stays visible until it is resolved. Correcting does not authorize deleting provenance. | MILPA-R-008 |
ASILO¶
| ID | Principle | Derived requirements |
|---|---|---|
| ASILO-P-01 | A model's output is an untrusted proposal until it crosses a boundary that validates it and authorizes it for whoever requests it. | ASILO-R-001, ASILO-R-002 |
| ASILO-P-02 | A compromised agent should not be able to do more harm than the party that invoked it. Its scope is declared before it operates and can be revoked. | ASILO-R-004 |
| ASILO-P-03 | Context is minimal, classified and time-limited. What does not enter the context cannot be leaked or confused. | ASILO-R-003, ASILO-R-006 |
| ASILO-P-04 | A person decides on critical effects and on ambiguity, not on every action. Human review is bounded so that it is real. | ASILO-R-005 |
| ASILO-P-05 | Resources have limits and the system knows how to stop without losing the evidence of what it did. | ASILO-R-007 |
| ASILO-P-06 | AI failures are measured and recorded; they are not declared eliminated. A control reduces a risk under tested conditions; the residual risk is named. | ASILO-R-008 |
The twelve statements come from the editions accepted on 28 September 2026. This current editorial presentation preserves their text and the provenance of the original documents. See the version history.