Published research on sovereign AI architecture, structural governance, deterministic state management, and personal AI infrastructure. All papers are open-access with permanent Zenodo DOIs.
Deployment Benchmarks and Artifact-Backed Evidence from 10 Months of AI State Management
Desai, P. (2026). Preprint. DOI: 10.5281/zenodo.18910362
Upgrades the SCD v4 deployment report into an artifact-backed systems paper. The underlying system, MirrorDNA, ran continuously from May 2025 through March 2026 on a single Apple M4 Mac mini with 24 GB unified memory. Benchmarks five aspects of the deployed state layer: event-store integrity, governance effectiveness, mutation integrity, cross-agent handoff envelopes, and context survivability under compaction pressure. Includes rerunnable measurement scripts, 10 benchmark tables, redacted sample data, protocol schemas, and a claim ledger mapping every headline number to its evidence source.
Desai, P. (2026). Preprint. DOI: 10.5281/zenodo.18888291
Introduces Personal AI Infrastructure (PAI) as a new computing paradigm in which individuals operate sovereign AI systems on hardware they own and govern. Presents MirrorDNA, a fully sovereign PAI system running on a Mac Mini M4 ($120/month total cost) that orchestrates 60 autonomous services, 85 LaunchAgent daemons, 259 operational scripts, and a 51,000-note knowledge vault without cloud dependency. Built over 10 months by a solo developer with AI pair-programming, MirrorDNA demonstrates that AI infrastructure with properties typically associated with institutional deployments can be constructed and operated as personal infrastructure.
A Human-Centered Architecture for Multi-AI Cognitive Extension
Desai, P. (2026). Preprint. DOI: 10.5281/zenodo.18712299
Presents an alternative to autonomy-first multi-AI architectures. The Orchestrator Model preserves human sovereignty while distributing computation across heterogeneous AI systems. Formalizes governance envelopes, model-agnostic identity continuity, and inter-model arbitration as first-class architectural primitives. Validated through system-level evidence from a 10-month operational deployment: 17,626 lines of control plane code, 2,386 commits across 73 repositories, a hash-chained audit trail with verified integrity, and an input security gate achieving 75% attack block rate with 0% false positives.
A Deterministic, Vendor-Independent Protocol for Persistent, Verifiable Agent State
Desai, P. (2026). Preprint. DOI: 10.5281/zenodo.18629531
A deterministic external state protocol that relocates agent memory from volatile model internals to cryptographically protected filesystem artifacts. Uses RFC 8785 JSON canonicalization for byte-identical serialization and SHA-256 integrity chains for tamper-evident state history. Demonstrates 100% determinism over 1,005 sequential state transitions, cross-vendor continuity between Gemini and Claude, and constitutional governance resistant to indirect prompt injection. 22 pages with formal proofs.
Percolation, Structure, and Greedy Routing in MirrorDNA
Desai, P. (2026). Preprint. DOI: 10.5281/zenodo.18629363
Studies a family of deterministic graphs defined by elementary number-theoretic rules: the prime-sum graph P_n and the power-of-two difference graph D_n. Demonstrates a sharp connectivity phase transition at k* = 19 with only 0.096% density. D_n is a strong small-world graph with clustering 28x random and path lengths only 1.44x random. Greedy routing achieves 100% delivery with 97.4% BFS-optimal paths. Includes NAF distance bounds, closed-form triangle count theorems, and automorphism group analysis. 13 pages with reproducible JSON evidence.
Separating Structural Authority Enforcement from Content Safety
Desai, P. (2025). Preprint. DOI: 10.5281/zenodo.18212082
Extends prior work on structural governance by introducing empirical evaluation and a layered architectural decomposition. Evaluation across 130 test cases shows that a governance-only layer achieves 70--100% refusal accuracy on authority-based attacks (prompt injection, autonomous execution, identity impersonation) while correctly passing all benign queries (0% false positive rate). Formalizes the governance pipeline as G . C . M, where G is a deterministic structural gate, C is a probabilistic content classifier, and M is the underlying language model.
Structural Enforcement Beyond Prompt Alignment
Desai, P. (2025). Preprint. DOI: 10.5281/zenodo.18212080
Argues that governance constraints for reflective AI systems -- those capable of maintaining state across interactions and reasoning about their own operation -- cannot be reliably achieved through prompt-level instructions alone. Presents a framework for structurally enforced governance through external wrapper architectures, deterministic state protocols, and cryptographic integrity verification. Validated through the Structured Contextual Distillation protocol, which demonstrated deterministic behavior across 1,005 sequential state transitions and successful cross-vendor continuity.
Hardware Specification Correction
Desai, P. (2025). DOI: 10.5281/zenodo.17926416
Corrects the deployment hardware specification in SCD v3.1 Section 6.1 from Mac mini with 16 GB RAM to Mac mini with 24 GB RAM. This correction does not affect any results, analyses, claims, or conclusions. The erratum has been incorporated into SCD v3.1.1.
This repository also contains essays and build logs in the articles/ directory covering sovereign AI architecture, multi-agent orchestration, infrastructure design, and operational lessons from building and running MirrorDNA.
@misc{desai2026mirrordna,
author = {Desai, Paul},
title = {Active Mirror Research Publications},
year = {2026},
publisher = {Zenodo},
url = {https://github.com/MirrorDNA-Reflection-Protocol/publications}
}Individual paper DOIs are listed in the table above. Please cite the specific paper relevant to your work.
Research papers are published under CC BY 4.0. Articles in the articles/ directory are published under the same license unless otherwise noted.
Published by Active Mirror -- Governed AI for Institutional Work.