research / papers
Research Papers
Published research from Cube Commons. Freely accessible.
All papers arise from the busCUBE and signCUBE development programs.
ANTS 2026 Late-Breaking Track — Submitted
Fleet Divergence Parameter Ψ: An Order Parameter for Cognitive Stigmergy in LLM Agent Systems
Introduces the Fleet Divergence Parameter Ψ, derived from Partial Information Decomposition, as an order parameter for cognitive stigmergy in LLM agent systems. Demonstrates that Ψ empirically subsumes the concurrent K* metric and provides a convergence proof for stigmergic coordination under exponential decay.
arXiv Preprint — In Preparation
Cognitive Stigmergy Across Federated Trust Boundaries: A Formal Model for AI Agent Coordination
A formal model of cognitive stigmergy — coordination through environmental traces rather than direct communication — for LLM agent systems operating across federated trust boundaries. Contributes a six-tuple digital pheromone definition, a typed taxonomy of six pheromone categories, a cross-boundary propagation model, a convergence proof, and an empirical analysis of a production 21-agent system. Survey of 24 LLM multi-agent frameworks reveals statistically significant bimodal convergence (Mann-Whitney p < 0.001, Cohen's d = 4.08).
arXiv Preprint — In Preparation
Intrinsic Access Control: A Sixth Access Control Model Class for AI Agent Systems
Formalizes Intrinsic Access Control (InAC) — a new access control model class defined by Subject-Enforcement Identity, where the entity that reads a security norm is the same entity that must enforce it. Proves four theorems including InAC Independence (irreducible to existing models) and the Substrate Theorem. Introduces InAC-Bench, a benchmark methodology producing aggregate resistance R(s) = 0.890 (cooperative) vs 0.446 (adversarial).
arXiv Preprint — In Preparation
The Passive Broker: A Third Category in Messaging Architecture Taxonomy for AI Agent Coordination Systems
Identifies and formally characterizes the passive broker as a third messaging architecture category — intermediated by a passive data structure (not a process). Survey of 24 independently-developed LLM coordination systems shows 67% convergence on passive broker patterns (p < 0.001, Cohen's d = 4.08). Provides formal taxonomy, decision boundary framework, queueing-theoretic scaling analysis, and anti-pattern reassessment.
arXiv Preprint — In Preparation
The Enforcement Location Principle: Why 40 Independent Research Groups Converge on Three-Layer Security for AI Agent Systems
Formalizes the Enforcement Location Principle (ELP) — the architectural framework for combining Intrinsic Access Control (InAC) with classical deterministic controls. Analyzes 40 independent research groups and derives the three-layer security pattern as a convergent design attractor.
NIST AI 100-4 RFI — Submitted March 2026
NIST CAISI RFI Response: AI Agent Security and Intrinsic Access Control
Formal response to the NIST AI 100-4 Collaborative AI System Integrity (CAISI) Request for Information. Presents the InAC model, ELP, and passive broker architecture as a combined framework for AI agent security in production deployments. Includes a companion addendum from the research team.
IETF Internet-Draft — draft-nicholls-dnssd-compute-00
DNS-SD for Computational Services
Proposes DNS-based Service Discovery (DNS-SD) extensions for advertising and discovering computational services — including AI agent endpoints and .cube namespace resolution — using existing DNS infrastructure. Companion to the cube-dns-sd CLI implementation under NLnet NGI Zero review.
For preprint access or collaboration inquiries: hello@cubecommons.org
Open research. Open source. Open future.