Translation
Translated from the Spanish source, which is the normative text. Status: draft pending human review. Differences are resolved in the Spanish text.
ASILO: analogies across disciplines¶
ASILO explains boundaries and security for systems built with artificial intelligence (AI) assistance. Comparing a computing boundary to a ship's compartment, or a review to a feedback loop, can help explain what a control aims to achieve. This page is for people who need to explain those decisions without starting with code.
The following correspondences are editorial adaptations. The sources describe their own mechanisms; they neither validate ASILO nor establish a direct genealogy of the methodology. The ASILO 0.1 requirements retain their scope, conditions and exceptions. This is an informative translation of the Spanish source.
Software protection and organisation¶
A code dependency indicates which part uses definitions from another. A privilege authorises particular operations. An execution boundary limits what a component can do while running. A drawing can show layers without demonstrating the other two properties; these references therefore need to be distinguished.
| Reference and source | Idea it helps explain | Explanatory adaptation in ASILO | Limit of the comparison |
|---|---|---|---|
| Protection rings: Schroeder and Saltzer, 1972. | Multics uses processor mechanisms to control privileges and crossings between rings. | Ask which actor may cross a boundary, through which mechanism and with what evidence. | Architectural rings drawn in ASILO are not processor privilege rings. Their position does not demonstrate isolation. |
| Ports and adapters: Cockburn, 2005. | Separate an application's interior from external devices and technologies through interaction contracts. | Explain why a contract can remain while its implementing mechanism changes. | The hexagon does not prescribe six parts. Code separation does not authorise actions or contain hostile code by itself. |
| Onion Architecture: Palermo, 2008. | Organise dependencies towards a core and place infrastructure at the periphery. | Show responsibilities and dependency direction in a declared model, as ASILO-R-002 requires where applicable. | Distance from the centre determines neither trust nor permissions. The requirement specifies no universal layer count. |
| Clean Architecture: Martin, 2012. | Distinguish internal policies and external mechanisms through the direction of code dependencies. | Explain what each layer needs to know and which changes are intended to stay local. | The circles are schematic. Meeting an import rule demonstrates neither authorisation of effects nor absence of errors. |
| Capability systems: Saltzer and Schroeder, 1975, section II.B. | A protected, unforgeable capability authorises access to an object. | Explain permissions bounded to resources and operations, with validity and revocation specified by the profile. | A written permission needs an effective mechanism. A capability is not a cryptographic zero-knowledge proof; restricting access does not demonstrate that property. |
| Defence in depth: NIST definition. | Combine people, technology and operations through complementary barriers. | Review contracts, authorisation, data and effects as different controls, with evidence for each. | Accumulating controls guarantees neither security nor independent failures. This definition does not prove a single military origin for ASILO. |
Containment, review and evidence¶
Containment means limiting how a failure spreads. Feedback means that a result or observation influences the system that produced it. Here it helps explain observing, reviewing and checking again. Self-similarity describes a resemblance of structure across scales; it is used here as an analogy, without attributing mathematical properties to the methodology.
| Reference and source | Idea it helps explain | Explanatory adaptation in ASILO | Limit of the comparison |
|---|---|---|---|
| Maritime compartmentalisation: IMO, damage stability. | Watertight subdivisions and functioning boundaries help prevent progressive flooding. | Identify what could propagate a failure and which boundary, control and check limit that path. | A separate folder or package is not a watertight bulkhead. Ship survival results do not transfer to software. |
| Electrical protection: Schneider Electric, ComPacT NSX guide. | Overload and short-circuit protection uses thresholds, delays and coordination between devices. | Explain a declared limit, what happens when it is exhausted and how stopping is checked, as in ASILO-R-007 where applicable. | A token budget is not electrical protection. Device curves, certifications and response times do not establish that software stops. |
| Immunology: NIGMS/NIH, 2024. | Recognition of self and nonself, and adaptive immune memory. | Distinguish provenance, treat external inputs with caution and preserve experience for reviewing controls. | A data schema is not a biological signature: shape validation verifies neither identity, intent nor truth. A record does not create automatic adaptation. |
| Cybernetics and regulation: Ashby, 1956, chapters 4, 5, 10 and 12. | Observe, compare with a desired condition, intervene and observe again. | Explain review of results, consumption and limits; check stopping or review actions in their environment. | Records do not stabilise a system by themselves. Feedback can be insufficient or unstable; automatic adjustment needs its own justification and checks. |
| Self-similarity geometry: Hutchinson, 1981. | The source formalises self-similar sets through families of contractions. | Recall questions about boundaries, permissions and evidence when reviewing an operation, a service and a system. | Repeating questions or patterns demonstrates neither that ASILO is a mathematical fractal nor that a part's security transfers to the whole. |
| Chain of custody: NIST definition and SP 800-86, 2006. | Document evidence movement and handling, with responsible people, dates and transfer purposes. | Record the result obtained, its revision and who received it, preserving provenance and changes. | A hash supports integrity comparison against a reference; it alone demonstrates neither authorship, a complete chain, truth nor legal admissibility. |
An example without code and with one assistant¶
A fictional team prepares a newsletter with invented data. A person asks an AI assistant to draft a summary; the task is to prepare text, and the person sends it after reviewing the final version.
Compartmentalisation helps consider which documents enter the task and which stay outside. Bounded permissions distinguish drafting from sending: producing text grants the assistant no sending authority. Feedback helps review errors and repeat the check on the correction. Chain of custody reminds the team to retain the source, reviewed version and sending decision.
The example produces reviewable records and decisions. No analogy demonstrates that the summary is true or that a control is implemented. A multi-agent system is not needed to ask these questions; the AI scope guide distinguishes the conditions of each situation.
To start, choose a reference that helps explain a project decision. Identify the actual control, its application condition, the observed result and what remains outside the check. Continue with layers and rings, capabilities and effects and evaluation.