Yellow Paper
The full technical specification: iCore architecture, personal learning, network cooperation, security, and implementation requirements.
Request accessEmail hello@quorumz.com to request a copy.
The HiNet library
Docs & research.
One place to understand HiNet, explore the science, and follow what comes next.
01 / Understand the project
The papers
Start with the idea. Go deeper into the product and its design.
Litepaper
The idea, the ownership model, and where HiNet is today. A concise introduction to the project.
Read paperWhitepaper
How iCores, iQuorums and iCorps work: the product, personal learning, privacy, and the network.
Read paperBrowse the library freely. Individual papers require an invited email. Request reading access ↗
02 / Explore the science
Research
The program, evidence and papers behind human-owned intelligence. Each work distinguishes results, assumptions and open questions.
Program, evidence & roadmap
See what we are working on, what the evidence supports, and what needs to happen next.
Research program & roadmap
Current research tracks, their status, next steps, and the criteria for becoming part of the implementation.
ExploreEvidence & next experiments
The September 14 evidence review: conditional proofs, measured gains, scaling limits, and proposed experiments.
ExploreWhat the Atomic Alignment research contributes
Separates our A2RL architecture and finite proof construction from the established methods they use. It identifies the original experimental evidence, corrects earlier numerical shorthand and states what remains unproved.
ExploreFoundations & research papers
Personal learning, human authority and cooperation between specialized models.
Atomic Alignment - Decentralized Intelligence Networks of Human Bound Specialists
Brings the five research papers together, distinguishing A2RL personal learning from independently learned recursive cooperation. It proves conditional learning and recursive-routing results, with explicit limits on privacy, network scale and comparisons with centralized AI.
Read paperPersonal Learning and Recursive Cooperation
Separates how an iCore learns its owner's preferences and skills through A2RL from how it learns to ask other iCores for help. It proves conditional bounds for answer quality, activated work and parallel response time, and shows why both learners must be evaluated together.
Read paperAtomic Alignment and Learning in Networks of Specialized Intelligence
Explains how an iCore learns from its owner, consolidates experience nightly and learns which peers to call. It proves conditional bounds on personal learning, controlled updates and cooperation, while distinguishing owner alignment from guaranteed benefit to everyone.
Read paperOwner Trained Mixtures of Specialized Intelligence
Shows how continuing owner feedback can create better personal specialists and how learned routing can preserve that benefit. Its proofs state the information and resource conditions needed for an advantage, including the limits of a network with a hundred million experts.
Read paperAtomic Alignment and Human Authority in Decentralized Policy Networks
Proves how independently enforced permissions can preserve human control and limit information exposure during cooperation. It also explains why remote stopping has limits, disclosed knowledge cannot reliably be recalled, and consent alone does not guarantee beneficial behavior.
Read paperFractal and Holographic Learning in Decentralized Policy Networks
Defines how smaller learning systems can compose recursively and what fractal and holographic learning would mean mathematically. It proves conditional results for cooperation, reconstruction and resource use, while identifying when a centralized system can match the network.
Read paperNatural Quorums and Learned Connectivity
Explains how permitted Telegram, WhatsApp and Slack relationships can help iCores discover useful collaborators. It proves conditions for reducing search costs and bounds the risks of missed expertise and revealing private relationships.
Read paperExperiments & research directions
Study designs, historical results and ideas still being tested.
The Latent Thought Layer
Proposes separating internal reasoning from language generation and allowing compatible iCores to exchange internal representations. The supporting calculations and simplified simulations do not establish the proposed benefits in trained language models, and a minimal linear graft on a small model failed.
Read paperThe Named-Concept Model
Proposes learning and reasoning with variable-sized concepts before translating them into language. It provides an architecture, prototype and evaluation criteria, but does not yet demonstrate better intelligence or efficiency in a trained language model.
Read paperE3 — Pre-Registration
Defines pass and fail criteria for separating reasoning from articulation and comparing internal-representation exchange with text communication. Its proposed accuracy, efficiency and privacy targets are research requirements rather than established capabilities.
Read paperE1 — Composition
Records a small three-specialist experiment where a merged model answered a mixed set of questions better than the best single specialist. Oracle routing performed substantially better still, revealing both the value of specialization and the limits of the tested weight merges.
Read paper03 / Watch & listen
A guided introduction to the research
Personal learning, recursive cooperation and the conditions for network advantage.
Read the transcript & sources