Milan Cheeks · Research

Private research.

Papers and preprints by Milan Cheeks: his private research, written as an independent researcher, and technical papers on the technologies he has built. Milan is enrolled in the M.S. in Computer Science program at the University of Colorado Boulder. His private research is conducted independently: no company sponsors or houses it. ORCID: 0009-0008-7507-0050 · Google Scholar · Corresponding email: mich3592@colorado.edu

Preprint · Technical Report · November 19, 2026 · Not yet peer reviewed · DOI: 10.5281/zenodo.24380531

Harmony: Media-Carried Workflow Signals for Distributed AI Agent Execution

AI agent systems are conventionally triggered through applications, APIs, or cloud consoles, while the media people already consume initiates nothing. Harmony embeds machine-readable workflow signals in the audio track of music and video and connects playback to approved agent work. The embedded cue carries a routing identity — an organization, a workflow, a channel, and an agent suite — rather than a fixed task, so a media asset authored once can address a task queue whose contents evolve over time, without re-encoding or redistributing the asset. The paper describes the complete prototype (encoder studio, signal inspector, reliability bench, listening node, runtime, and template executor), analyzes its security and privacy posture including its known weaknesses, and specifies a five-part validation program for converting the internally demonstrated loop into a measured delivery channel. The system described is patent pending (USPTO Provisional No. 63/587,817, converted).

Read the paper (PDF)

Cite as: Cheeks, M. (2026). Harmony: Media-Carried Workflow Signals for Distributed AI Agent Execution. Zenodo. https://doi.org/10.5281/zenodo.24380531

Preprint · Architecture and Position Paper · November 19, 2026 · Not yet peer reviewed · DOI: 10.5281/zenodo.24384875

BLOOD: A Biological Root of Trust and Dynamic Memory Architecture for Advanced Computing Platforms

Every recovery mechanism in modern computing is itself software or firmware, inside the same attack surface it defends. BLOOD (Biological and Lightweight Organics Optimized for Dynamic memory systems) integrates engineered exosomes with optogenetic control to form a vein-like biological network inside computing platforms and humanoid robots. The paper presents BLOOD as a dynamic memory and signaling substrate and, as its central thesis, as a biological root of trust: a layer outside the digital attack surface that holds a known-good recovery state and can force a compromised platform back to it — a hack-proof factory reset. It states the architecture, the threat model and its limits, and a staged validation program, and it is explicit that BLOOD is at the architecture stage with no bench results yet. BLOOD is a SELF Labs Inc technology, patent pending (U.S. Patent Application No. 18/901,437); this paper is authored by its founder and published through his research channel.

Read the paper (PDF)

Cite as: Cheeks, M. (2026). BLOOD: A Biological Root of Trust and Dynamic Memory Architecture for Advanced Computing Platforms. Zenodo. https://doi.org/10.5281/zenodo.24384875

Preprint · Research Manuscript · November 19, 2026 · Not yet peer reviewed · DOI: 10.5281/zenodo.24385052

PRESENCE: Adaptive Intelligence Without Continuous Model Retraining

Foundation models can reason about unfamiliar tasks, but a successful interaction does not automatically produce persistent operational knowledge. PRESENCE (Persistent Role-Adaptive, Environment-Native, Self-Evolving Cognitive Engine) proposes an architecture for continuous system-level behavioral adaptation while the foundation model remains frozen: an evolving external cognitive state — environment representations, contextual memory, role definitions, validated skills, operational histories, and quality estimates — driven by a cycle that observes, interprets, retrieves, plans, acts, verifies, and consolidates experience. The paper distinguishes parameter learning from externally mediated operational learning and specifies controlled comparisons against memoryless, retrieval-based, experience-memory, and skill-based agents. It is a pre-empirical manuscript: the architecture and evaluation protocol are proposed, and no completed experiments are reported.

Read the paper (PDF)

Cite as: Cheeks, M. (2026). PRESENCE: Adaptive Intelligence Without Continuous Model Retraining. Zenodo. https://doi.org/10.5281/zenodo.24385052

Preprint · Theoretical Economics Working Paper · November 19, 2026 · Not yet peer reviewed · DOI: 10.5281/zenodo.24385420

The Deflationary Effects of Generative AI on Digital Goods and Services

Generative AI may transform the economics of digital production by reducing the cost of cognitive tasks, expanding productive capacity, lowering entry barriers, and enabling consumers to produce substitutes for commercially supplied services. This paper proposes a conditional theory of AI-induced relative-price deflation across five channels — declining unit production costs, increased entry, competitive repricing, consumer self-production, and changing vertical market structure — with a partial-equilibrium model showing that prices fall when the supply effect exceeds AI-induced demand expansion, and a complementary model of downstream commoditization alongside upstream concentration. It develops falsifiable hypotheses and an empirical research program across software, creative services, customer support, and professional services, and argues that sectoral price declines need not imply economy-wide deflation. Theoretical working paper; no original causal price estimates are claimed. JEL: D21, D43, L11, L86, O33, E31, J24.

Read the paper (PDF)

Cite as: Cheeks, M. (2026). The Deflationary Effects of Generative AI on Digital Goods and Services. Zenodo. https://doi.org/10.5281/zenodo.24385420