Architectural reference · concept DOI 10.5281/zenodo.20117025
MILO (Modular Intelligent Learning Orchestrator) is an adaptive AI orchestration architecture engineered around one structural commitment: human authority is preserved as an architectural property, not a deployment policy. The system composes independent learning modules behind an audit-first command bus; consequential outputs are deferred to an authorization layer with latency-aware semantics; failure of any single learning component reroutes traffic to a peer without operator intervention. The reference described here, together with five companion articles, establishes a public, reproducible, and citable basis for adaptive AI systems whose authority structure can be inspected and verified.
This page is the persistent public landing for that reference. Each linked artifact (article, DOI, repository commit, patent filing) is the canonical record; the page itself is not authoritative when in tension with the linked record.
Each article has an independent DOI, is mirrored as HTML, plain text, Markdown, and PDF, and is licensed CC BY 4.0. Together they constitute the published, citable basis for MILO. A BibTeX collection is at the end of this page.
Why module independence has to be designed in at the level of the message bus and audit substrate, not retrofitted at the deployment layer. Defines a thermodynamic notion of independence based on entropy of inter-module signal traffic.
An adaptive authentication discipline for industrial control and human-in-the-loop AI environments where authorization decisions arrive on millisecond timescales. Treats latency as an input to the authentication function rather than a constraint on it.
How human authority can be represented as an architectural property — a measurable invariant — rather than as a policy layer that can be removed or downgraded. The conceptual core of the MILO commitment to human-in-the-loop AI.
Engineering constraints — audit-first command flow, persist-before-deliver, bounded recovery, reviewed-outcome learning, preserved operator authority, and three more — that together yield a system designed for viability under unforeseen operational futures rather than predictive accuracy on familiar ones.
A treatment of resilience as the system property that emerges when independence (Article 01), latency-aware authentication (02), supervisory primacy (03), and the structural principles (04) compose. Includes a working definition of operational viability under tail-event load.
The reference describes the architecture in terms of seven primitives. Each is independently realizable and independently testable; together they realize the published commitment to adaptive AI under human authority.
MILO is grounded in established standards rather than asserted as a novel framework de novo. The reference is consistent with, and cites where appropriate, the following:
Lineage: the architecture draws on Beer (Viable System Model), Ashby (requisite variety), Shannon (information), Lyapunov (stability), Hollnagel (resilience engineering), Taleb (antifragility under tail load), and Clauset (heavy-tailed event statistics).
Cite the concept DOI for the architecture as a whole; cite per-article DOIs when referring to specific findings. The repository commit hash may be cited when referring to a particular implementation snapshot.
@misc{milo_architectural_reference,
author = {Flores Montano, Jorge Enrique},
title = {{MILO} -- {Modular Intelligent Learning Orchestrator}: A Public Architectural Reference},
year = {2026},
publisher = {Zenodo},
doi = {10.5281/zenodo.20117025},
url = {https://www.jmautomated.com/architectural-reference},
note = {Patent pending, USPTO Provisional 63/993,825 (confidential during pendency). License CC BY 4.0.}
}
@article{milo_article_01_independence,
author = {Flores Montano, Jorge Enrique},
title = {Independence as an Architectural Property},
year = {2026}, publisher = {Zenodo},
doi = {10.5281/zenodo.20117647}
}
@article{milo_article_02_latency_auth,
author = {Flores Montano, Jorge Enrique},
title = {Latency-Aware Authentication},
year = {2026}, publisher = {Zenodo},
doi = {10.5281/zenodo.20117651}
}
@article{milo_article_03_supervisory_primacy,
author = {Flores Montano, Jorge Enrique},
title = {Supervisory Primacy},
year = {2026}, publisher = {Zenodo},
doi = {10.5281/zenodo.20117662}
}
@article{milo_article_04_structural_principles,
author = {Flores Montano, Jorge Enrique},
title = {Eight Structural Principles for Adaptive {AI}},
year = {2026}, publisher = {Zenodo},
doi = {10.5281/zenodo.20117703}
}
@article{milo_article_05_adaptive_resilience,
author = {Flores Montano, Jorge Enrique},
title = {Adaptive Resilience},
year = {2026}, publisher = {Zenodo},
doi = {10.5281/zenodo.20117716}
}
A guidance page is also available at /cite, and machine-readable formats (txt, markdown, RSS, sitemap, llms.txt, well-known/ai.txt) are linked from /formats.
If you arrived here from outside the research literature, here is the same statement in plain terms. Most AI systems are deployed under the assumption that the operator will catch mistakes after the fact. MILO is built on the opposite assumption: that mistakes should be caught before the fact, that every action should leave a record, and that the human responsible for a decision should never be displaced by the system that recommended it. The five companion articles each take one piece of that commitment and develop it carefully enough to be reviewed, cited, and (where applicable) refuted. They are open to read in full, in any format, at no cost.
If you are a researcher, the concept DOI is the place to start. If you are a journalist or a policy reader, the supervisory-primacy article is the conceptual core. If you are a practitioner, the structural-principles article is the implementation contract.