Translation

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

AI, assistants and agents: methodology scope¶

ASILO and SCAFFOLD include work with artificial intelligence (AI). They do not require a multi-agent system. ASILO bounds systems built with AI assistance and model uses that activate its requirements; SCAFFOLD manages software projects built with AI, from framing to delivery. MILPA preserves knowledge and provenance in both contexts.

Terms to distinguish¶

Term	Meaning here
AI	A field and techniques used to produce predictions, content or decisions; it does not identify an architecture by itself.
Model	A component processing input and producing output; using a model does not imply several agents.
AI assistant	A tool supporting a person in a task, such as writing, analysis or programming.
Agent with tools	A component with identity and scope that can act through tools or access data. ASILO-R-004 conditions apply.
Multi-agent system	A system coordinating several agents. It is one application case, not an entry requirement of ASILO or SCAFFOLD.

Three situations, not a maturity ladder¶

flowchart TD
  accTitle: AI scope and application conditions
  accDescr: The branches are situations, not maturity levels. Requirements retain their own conditions.
  T["Task or system with AI assistance"] -->|model producing content| M["Interaction with a model"]
  T -->|construction support| S["One person and one assistant"]
  T -->|additional coordination| V["Several people or assistants"]
  M --> C["ASILO: requirements according to their conditions"]
  S --> C
  V --> C
  S --> E["SCAFFOLD: software projects with people and AI assistants"]
  V -->|when building software| E
  C --> R["MILPA: project sources, authority and evidence"]
  E --> R

SVG diagram: AI scope and application conditions · Editable Mermaid source.

Arrows separate situations and related methods. They require no sequence and do not mean that several AIs provide greater certainty. A single assistant can build software; a model can produce content without using tools; several agents introduce coordination risks. Each project records its case.

Application conditions that remain in force¶

Requirement	Relevant condition
ASILO-R-001	Effect points fed by a model. Shape validation proves neither truth nor authorisation.
ASILO-R-002	More than one module or package in a system declaring ASILO; no layer count is imposed.
ASILO-R-008	Before claiming fitness publicly or to a customer, and on each relevant change of model or configuration. Failures between agents apply when there is more than one.
SCAFFOLD-R-002	A change touches several owners’ zones; this does not require creating several agents.
SCAFFOLD-R-005	The assistant changes within a batch.
SCAFFOLD-R-016	One batch depends on another, even if the same person runs both.

This table is guidance: complete records govern scope, actor, evidence, exceptions and residual risk. A summary, diagram or guide does not extend accepted obligations.

ASILO: building with AI and systems that contain AI¶

The table compares two situations: (a) the system contains a model or an agent; (b) the system is only built with an AI assistant and contains none. Each quotation comes from the scope or the applicability condition of the record.

Requirement	Quotation from the record	(a) Contains a model or agent	(b) Only built with an assistant; contains none
ASILO-R-001	Condition: "Whenever there is at least one effect point fed by a model."	Yes, if a model feeds at least one effect point; if it only shows text to a person, the exception is recorded.	Not in the built system. Effects of the assistant during construction: depends on the profile / see the record.
ASILO-R-002	Condition: "Systems with more than one module or package that declare that they apply ASILO."	Yes, with more than one module or package and ASILO declared.	Same as (a): the condition does not depend on a model being present.
ASILO-R-003	Scope: "Each task in which an agent or model receives context"; condition: "with access to internal sources, personal data or persistent memory".	Yes, in tasks of the model or agent with that access.	Depends on the profile / see the record.
ASILO-R-004	Scope: "Agents, subagents and tools that execute actions."; condition: "Any agent with access to tools or data."	Yes, for each agent with access to tools or data.	Yes, if the assistant acts as an agent with access to tools or data; otherwise, no.
ASILO-R-005	Condition: "When an action belongs to the catalog of critical effects or the system cannot decide with its rules."	Yes, when the condition is met.	Yes, when the condition is met; it does not depend on a model being present.
ASILO-R-006	Scope: "Personal data, data of minors, credentials and confidential data"; condition: "Whenever the system processes any of that data."	Yes, if the system processes that data.	Yes, if the system processes that data; it does not depend on a model being present.
ASILO-R-007	Scope: "Agent runs and model calls with cost or effect."; condition: "Any automated run with cost or effects."	Yes, for runs and calls with cost or effect.	Depends on the profile / see the record.
ASILO-R-008	Scope: "Systems that use models to produce content, decisions or actions […]"; condition: "Before claiming publicly […] that the system is fit".	Yes; failures between agents only if there is more than one. Marked internal prototypes are the record's exception.	No: the scope is systems that use models.

This table is a reading aid; where it differs from the record, the record prevails.

Small examples¶

A person uses AI to write code: SCAFFOLD organises zones and batches; ASILO bounds relevant context and actions; MILPA records sources and decisions. The final product may not use AI.
An application queries a model for a summary: ASILO guides contracts, data and evaluation under applicable conditions; SCAFFOLD describes construction when it uses AI assistance.
A team coordinates several assistants: the same distinctions remain, with handoffs, dependencies and failures between agents reviewed where relevant. Quantity does not replace evidence.