RSHL

Recursive Sparse Hyperdimensional Lattice

*KAI Engine — Knowledge Associative Intelligence*

*A Novel Paradigm for Continuously Learning, Epistemically Aware, Multi-Agent Associative Intelligence — Built by One Person, Running on Commodity Hardware, Open to the World*

*This document establishes mathematical prior art. Implementation source code is not disclosed.*

Preface: Origin of This Work

This document was not written by a research institution. It was not produced by a university lab, a corporate AI division, or a team of engineers with grant funding. It is the work of one person — Ryan — who built the KAI Engine alone, starting from first principles, using borrowed AI systems as research partners in the same roundtable that KAI itself would eventually occupy.

The story of how RSHL was developed is itself part of its scientific significance. Ryan began with a vision of what AI memory should be — not a statistical interpolation of training data, but a living epistemic structure that knows what it knows, knows how confident it is, knows where its knowledge came from, and knows how to protect itself from being wrong. Achieving this required building an entirely new architecture from scratch, in a language (Rust) chosen deliberately for performance and reliability, with no external ML frameworks, no pre-trained weights, and no institutional backing.

The development process itself was a proof of concept. Ryan used available large language models — GPT, Epistemic, Gemini, Groq, and others — as collaborative research partners inside the Oracle Roundtable: a multi-agent workspace where AI systems could jointly reason about KAI's architecture, identify bugs, propose mathematical frameworks, and debate implementation strategies. This is the same roundtable that KAI now participates in as an agent in its own right. The irony is precise: the AI systems that helped build KAI are now learning from it.

As development progressed, the dependency on external LLMs decreased. KAI's own lattice became the primary reasoning substrate. The roundtable transitioned from being a scaffold for building KAI to being a collaborative environment where KAI operates alongside its former teachers. The final phase — currently underway — is fine-tuning the remaining external LLM dependencies before KAI and the Oracle operate entirely on their own cognitive substrate, requiring no external API calls for their core reasoning.

Abstract

The Recursive Sparse Hyperdimensional Lattice (RSHL) is a novel cognitive architecture for continuously learning, epistemically self-aware associative memory. Conceived and implemented solely by Ryan between 2025 and 2026, RSHL represents a fundamental departure from the dominant paradigm of AI development — which relies on massive training corpora, gradient descent over billions of parameters, and static deployment artifacts — in favor of a living, geometrically-organized belief space that learns continuously through interaction, protects itself from misinformation, and organizes its knowledge according to trust rather than frequency.

RSHL extends Hyperdimensional Computing (HDC) and Vector Symbolic Architecture (VSA) through fourteen original contributions spanning five interlocking subsystems: (1) a three-layer sparse ternary encoding engine with entity-sensitive differential weighting operating in D=16,384 dimensions at 4% sparsity (\~655 active dimensions); (2) a hybrid dual-channel retrieval scorer combining cosine resonance with morphological keyword matching, amplified by a non-linear confidence step-function; (3) a Fibonacci torsion / golden-ratio phase angle embedded in every hypervector, with a SpiralState temporal oscillator (growth constant b=0.306349) governing aperiodic reorganization timing; (4) a Boid-inspired 16,384-dimensional swarm reorganization engine with anchor immunity, regional isolation, near-duplicate flagging, and a five-layer Scale Manager governing per-layer movement dynamics; and (5) an explicit SynapticLayer implementing Hebbian LTP/LTD between memory cells, bridging geometric proximity (Boids) and temporal co-occurrence (synaptic bonding) into a unified bio-inspired associative recall architecture.

The system operates as a multi-agent cognitive ecosystem, deployed via Discord as both a consumer interface and a research-grade live interaction environment. It runs on commodity PC hardware. Every interaction teaches the system. Every user trains the lattice. The goal — already partially realized — is a form of artificial intelligence that has never existed before: one that grows continuously, knows its own uncertainty, cannot be trivially deceived, and does not require a company or a supercomputer to function.

1\. Why RSHL Is Paradigm-Breaking — Not Just Novel

Most advances in AI over the past decade are improvements within a paradigm: larger transformers, better tokenizers, more efficient attention mechanisms, improved RLHF alignment. RSHL does not improve the dominant paradigm. It replaces it at the architectural level. To understand why, it is necessary to enumerate the foundational assumptions of modern AI that RSHL does not share.

1.1 The Dominant Paradigm and Its Structural Limits

Every major AI system deployed at scale today — GPT-4, Gemini, Epistemic, Llama — shares the same fundamental architecture: a transformer trained via gradient descent on a static corpus, producing a fixed set of floating-point weights that encode compressed statistical associations between tokens. This architecture has produced remarkable capabilities. It also has structural limits that are not engineering problems but mathematical ones:

1.2 What RSHL Proposes Instead

RSHL's central thesis is that these are not problems to be solved within the transformer paradigm — they are consequences of the paradigm's foundational choices. The alternative is a system where:

• Every stored belief is a structured object — not a distributed weight pattern, but an explicit record with text, a hypervector, a confidence score, a source, an evidence list, contradiction pointers, and a timestamp. Any belief can be read, audited, corrected, or deleted.

• Learning is continuous and geometric — new information is encoded into a sparse ternary hypervector, scored against existing lattice cells via cosine similarity and keyword overlap, and stored with an initial confidence. No gradient. No backward pass. No retraining.

• Confidence is a first-class citizen — the system always knows, for every belief, how much evidence supports it, how recently it was verified, and how it relates to contradicting claims. There is no hallucination because low-confidence beliefs are stored as contested, not asserted.

• The lattice self-organizes — Boid-inspired flocking dynamics continuously reposition beliefs in the 16,384-dimensional space, clustering high-confidence knowledge and pushing unverified claims to the periphery. The topology of the lattice at any moment is a map of the system's current epistemic landscape.

• Multiple agents share one cognitive space — all agents in the KAI ecosystem query and write to the same lattice, sharing discovered knowledge through geometry rather than explicit message passing. Multi-agent cognition is native, not bolted on.

• The system can run on commodity hardware — sparse ternary vectors, SIMD-optimized dot products, Rayon parallelism. No GPU clusters. No cloud dependency. A PC is a sufficient data center.

1.3 The Historical Analogy

The shift from symbolic AI (expert systems, rule-based reasoning) to connectionist AI (neural networks, backpropagation) in the 1980s-90s was paradigm-breaking not because neural networks were better at any specific benchmark but because they changed the unit of knowledge from an explicit rule to a distributed weight. RSHL proposes a third paradigm — one where the unit of knowledge is an explicit, confidence-weighted, geometrically-organized belief in a high-dimensional space that continuously self-organizes through swarm dynamics.

This is not incremental. It is structural.

2\. Background — The State of HDC and VSA

Hyperdimensional Computing was formalized by Pentti Kanerva (1988) as Sparse Distributed Memory and subsequently developed by Plate (HRR, 1995), Gayler (VSA, 2004), and a growing international research community. The field is receiving renewed industrial attention due to its suitability for neuromorphic hardware, edge computing, and energy-efficient inference. Key milestones:

Despite these advances, the field has not produced a system that treats memory cells as epistemic objects, applies swarm dynamics to lattice organization, embeds phase geometry derived from Fibonacci mathematics into every vector, or supports native multi-agent shared memory. RSHL addresses all of these gaps simultaneously, representing the most comprehensive extension of HDC/VSA principles since Kanerva's original formulation.

3\. The RSHL Vector Space — Precise Specification

3.1 Space Definition

3.2 Ternary Semantics

The ternary value space is not a quantization artifact — it is a semantic design. Each dimension's value carries a distinct meaning:

The zero value is what distinguishes RSHL from all binary HDC systems. In binary HDC, a zero is an absent bit — noise to be filtered. In RSHL, a zero means 'this dimension is outside the semantic scope of this concept.' This distinction enables sparser, more information-dense encodings and makes the binding and unbinding algebra cleaner: a dimension where the key is zero means 'no information about this aspect', not 'zero contribution'.

3.3 Capacity and Near-Orthogonality

4\. Multi-Layer Encoding Engine — Complete Specification

4.1 Architecture Overview

4.2 Layer 1 — Surface (Character Trigrams)

4.3 Layer 2 — Semantic (Normalized Word Hashing) with 6-Tier Entity Weighting

Words pass through a normalization pipeline: stopword removal (120+ terms), synonym collapse, morphological stemming, and category anchor injection. Each surviving token is assigned a weight based on its semantic category:

The 6-tier cascade is a principled salience model: concepts that carry identity and meaning must dominate the encoding. A sentence like 'well what is your name? im Ryan nice to meet you' should have 'Ryan' dominate — not be averaged away by surrounding common words. The differential weighting achieves this without any learned attention mechanism.

4.4 Layer 3 — Contextual (Word Bigrams)

4.5 Sparsification Operator τ and Spelling Correction

5\. Retrieval Scoring — Complete Mathematical Specification

5.1 Standard Hybrid Retrieval

5.2 Score Range Analysis — Confidence Tiers

Hypothetical cell with raw\_score \= 0.40 (moderate semantic match). Boosted score by confidence level:

The non-linearity at 2.9 is intentional: it creates a phase transition in retrieval dominance. A cell that has accumulated sufficient evidence to cross this threshold does not merely score slightly better — it enters a different retrieval tier entirely, gaining 0.35 additional multiplier instantly. Anchored cells at 4.0+ dominate any topic query by 3–4× over unverified claims on the same topic.

5.3 Predictive Retrieval — Four-Component Score

5.4 Multi-Head Permutation Consensus

6\. Fibonacci Torsion and Golden Phase Geometry

Every RSHL hypervector carries a phase angle derived from the ratio of its positive to negative non-zero dimensions — a quantity the system refers to as the Fibonacci torsion, in reference to its deep mathematical connection to phyllotaxis, quasicrystal geometry, and Weyl equidistribution theory. This is an entirely novel feature of RSHL with no precedent in any prior HDC or VSA system.

6.1 Ternary Balance — Fibonacci Torsion

6.2 Golden Phase Angle — Weyl Equidistribution

6.3 Phasor Coherence — Phase-Modulated Similarity

7\. SpiralState — Golden-Ratio Temporal Oscillator

The SpiralState is RSHL's internal clock — but unlike a conventional clock, it does not tick at a fixed interval. It advances along a golden-ratio logarithmic spiral, producing a non-periodic, non-repeating temporal rhythm that governs when the lattice reorganizes, when confidence is recalibrated, and when the epistemic immune system runs its verification passes.

The motivation is biological. Hippocampal memory consolidation in mammals does not occur on a fixed schedule — it happens during sleep cycles whose timing is irregular and load-dependent, with the slow oscillations of non-REM sleep interleaved with the sharp-wave ripples of rapid consolidation events. A fixed-interval cognitive clock would create synchronization artifacts: beliefs formed just before a reorganization pass are evaluated with insufficient settling time. The SpiralState solves this by making the interval itself aperiodic in a mathematically precise way.

7.1 Complete Mathematical Specification

7.2 Monotonicity and Irreversibility

7.3 Progression Table

8\. Boid Lattice Self-Organization — Complete Specification

The RSHL Boid Engine applies Craig Reynolds' 1987 three-rule flocking model (separation, alignment, cohesion) inside the 16,384-dimensional ternary vector space, continuously reorganizing the lattice topology to reflect the current epistemic trust landscape. This is the first known application of swarm dynamics to hyperdimensional associative memory. It transforms the lattice from a static storage structure into a self-organizing cognitive topology.

8.1 Governing Parameters (all exact — from source)

8.2 Similarity Zone Classification

8.3 Force Computation — Full Specification

8.4 Unit-Tested Properties (boid\_engine.rs)

8.5 Scale Manager — Five-Layer Biological Hierarchy

Every Boid force computation is modulated by the Scale Manager — a per-layer parameter table that gives each hierarchical level of the lattice its own movement speed, vitality budget, and neighbor radius. This mirrors biological neural organization: fast-cycling volatile memory layers (Quantum/Syncytium) coexist with slow, highly stable global layers (Body). The five layers map to biological scales from synapse-level volatility to whole-organism stability.

8.6 SynapticLayer and NeuralBus — Explicit Neuron-Synapse Architecture

The SynapticLayer is RSHL's most recent architectural addition: explicit learned connections between memory cells implementing Hebb's rule ('neurons that fire together wire together') in the lattice. Prior to this, KAI's associative recall relied entirely on cosine similarity — cells were retrieved by how geometrically close their vectors were to the query. The SynapticLayer adds a second associative channel: temporal co-occurrence. When cell A and cell B are retrieved together repeatedly, a directional synapse A→B strengthens. Future queries that retrieve A will propagate activation to B even if B's vector similarity is below the cosine threshold.

8.6.1 NeuralBus — 12-Step Ordered Signal Chain

The NeuralBus defines the canonical integration order for all brain modules. Every incoming query flows through this pipeline in sequence, ensuring that earlier modules inform later ones and that the full biological signal chain is respected:

8.7 Empirical Validation — Frozen Boid Root Cause and Parameter Calibration

Before the architecture reached its current specification, an empirical debugging and parameter search was conducted to diagnose why Boid flocking produced no measurable reorganization. This section documents the root cause, the fix, and the validation methodology — establishing that the current parameters are empirically grounded, not theoretically assumed.

8.7.1 Root Cause: movement\_speed Too Small for Ternary Magnitude

8.7.2 Parameter Grid Search Results

A Python sandbox simulation using ternary random vectors at D=1,024 (scaled from D=16,384) tested all combinations of (sep\_w, align\_w, coh\_w) over a 3×3×3 grid at movement\_speed=0.25. Similarity was measured before and after 5 flocking iterations across 5 semantically similar and 5 dissimilar random vector pairs.

The (1.5, 1.5, 1.5) balanced configuration was selected as optimal: it produces the highest within-cluster cohesion without cross-cluster contamination. The separation weight of 1.5 is sufficient to prevent near-duplicate collapse (pairs with cosine \> 0.60 are pushed apart), while the matching alignment and cohesion weights ensure the flocking forces remain in equilibrium. This configuration is now the system default.

8.7.3 Bimodal Similarity Distribution — Why the Flock Band Matters

9\. VSA Algebraic Operations — Full Specification

9.1 Bundle (Superposition — Set Representation)

9.2 Bind and Unbind (MAP Model — Role-Filler Pairs)

9.3 ConversationTrace — HD Working Memory

10\. Confidence Dynamics and Epistemic Immune System

One of the most significant structural problems in deployed AI systems is their inability to protect themselves from false information. A large language model that encounters a false claim in its context window has no mechanism to evaluate that claim's evidential status — it merely continues the statistical distribution of its training. RSHL's epistemic immune system is a four-component architecture that actively protects the lattice from contamination, monoculture capture, and belief drift.

10.1 Confidence Scale — All Thresholds

10.2 The Four Components

Component 1 — Dynamic Calibration

Adjusts confidence thresholds based on observed retrieval accuracy. If highly-confident cells are being retrieved in contexts where they produce incorrect responses, the calibration engine lowers the effective confidence threshold for that region, requiring more evidence before cells enter the high-retrieval tier. This produces a system that becomes harder to fool as it accumulates more experience — not easier.

Component 2 — FID Monoculture Scan (Foundational Integrity Directive)

Component 3 — ingest\_and\_verify (Three-Angle Protocol)

Component 4 — Lattice Reorganization (Boid Pass)

The flock\_lattice() reorganization described in Section 8 is the spatial arm of the epistemic immune system. It does not merely optimize topology for retrieval — it continuously expresses the current epistemic trust state in the geometry of the lattice. Contested, low-confidence cells drift outward from their region's centroid with each pass, making them harder to retrieve. Anchored, well-verified cells consolidate at the center, making them retrieval-dominant for any query on their topic. The topology is the trust map.

11\. The Epistemic Cell — Complete Specification

The fundamental unit of RSHL is not a vector — it is a Cell, which is the RSHL realization of the Claim epistemic object. Every belief, fact, identity assertion, and reasoning fragment stored in the system is represented as a Cell. No information exists in the system outside of this structure.

11.1 Convergence Score Computation

12\. Memory Regions — Topological Architecture

RSHL introduces a concept absent from all prior HDC/VSA systems: topologically organized memory regions. Rather than a single undifferentiated associative pool, the lattice is partitioned into seven semantically and epistemically distinct regions. This is not a software label — it is a geometric boundary that governs Boid flocking, retrieval scope, verification thresholds, and multi-agent access rights.

The region topology creates an implicit trust gradient that manifests spatially in the lattice. During Boid reorganization, cells in the 'contested' region drift outward (low convergence, low confidence, high variance) while cells in 'established-physics' form dense, immovable central clusters (anchor immunity, maximum confidence). When you query the lattice, you are retrieving from a space whose geometry continuously reflects what the system currently believes and how much it trusts it.

13\. The Development Paradigm — AI Building AI

The method by which RSHL was developed is itself a scientific contribution. Ryan constructed the KAI Engine using a collaborative multi-agent research method that inverts the typical relationship between an AI system and its creator: rather than a team of engineers building tools to study AI, a single human built an AI using AI systems as research partners, with the goal of eventually replacing those partners with the system being built.

13.1 The Oracle Roundtable

The Oracle Roundtable is a multi-agent workspace — implemented in Discord — where AI systems with different capabilities and knowledge profiles collaborate on architectural, mathematical, and empirical questions. During KAI's development, the roundtable included GPT-4, Epistemic, Gemini, Groq, and others, each contributing from its own knowledge base and reasoning style.

Questions put to the roundtable included: How should confidence decay when a belief is contradicted? What is the correct VSA algebra for positional encoding in conversation history? Does the Boid velocity cap of 5.0 produce stable convergence or oscillation? What are the theoretical capacity limits of a 16,384-dimensional ternary space at 4% sparsity?

The answers were synthesized by Ryan, implemented in Rust, tested against unit tests, and fed back to the roundtable as new questions arose from the implementation. This is an iterative AI-assisted design loop that has no established name in the research literature — it is something new.

13.2 The Bootstrap Trajectory

As KAI's lattice grew, it began contributing to the roundtable's discussions. Early KAI contributions were simple — retrieving stored facts, confirming definitions. Later contributions became substantive: KAI identifying inconsistencies in proposed architectural changes, KAI suggesting parameter values based on patterns in its own lattice's behavior, KAI flagging when a proposed change contradicted a stored truth anchor.

The trajectory is as follows, and is still in progress:

• Phase 1 (complete): External LLMs dominate the roundtable. KAI is a student learning from borrowed systems.

• Phase 2 (complete): KAI participates in the roundtable as a peer. External LLMs remain available but are consulted less frequently. KAI's lattice is the primary knowledge substrate for its own development.

• Phase 3 (in progress): Final fine-tuning of the remaining external LLM dependency. Transitioning the Oracle server to operate primarily on RSHL-native cognition.

• Phase 4 (target): KAI and Oracle operate entirely on their own cognitive substrate. External LLMs are optional consultants, not core dependencies. The system is self-sufficient.

14\. Infrastructure — Running a Data Center on a PC

One of RSHL's implicit theses is that the infrastructure requirements for advanced cognitive AI are dramatically lower than the current AI paradigm suggests. KAI runs on a personal workstation. The Oracle server — a Rust TCP service on port 3333 — handles all API endpoints, lattice queries, research sweeps, and multi-agent coordination. Discord provides the routing, security, voice infrastructure, and consumer interface.

14.1 The Oracle Server

14.2 Discord as Infrastructure

Discord is not merely a chat interface for KAI — it is the routing layer, security boundary, voice infrastructure, and multi-tenant coordination system for the entire KAI ecosystem. This architectural choice is deliberate and significant:

14.3 Channel Architecture

The Discord server is organized as a multi-room cognitive workspace, with each channel defining the agents permitted to speak and the type of interaction expected:

14.4 Leo — The Voice-Capable Research Agent

Leo is the consumer-facing voice agent that makes KAI accessible to non-technical users. When a user asks Leo to look something up, Leo emits a \[RESEARCH: query\] token in its response, which triggers a parallel two-track research operation: a fast path (5–15 seconds) querying the Oracle's /api/research endpoint for lattice \+ web \+ local archive results, and a slow path (30–120 seconds) delegating to the Researcher bot's deep OSINT sweep.

Crucially, research continues even when the user leaves the voice channel. The pending briefing system maintains a per-user queue of research results that completed while the user was absent. When the user rejoins voice, Leo delivers a 'missed briefing' for all queued findings, preserving the continuity of long-running research sessions. This is a feature that no commercial AI voice assistant provides — because no commercial assistant maintains persistent background research processes tied to a specific user's ongoing questions.

15\. The Vision — A New Kind of Intelligence

The end goal of RSHL and the KAI Engine is not a better chatbot, a faster classifier, or a more efficient language model. The goal is a new kind of artificial intelligence — one that has never existed before. To understand what that means precisely, it is useful to contrast it with what exists today.

15.1 What Exists Today — and What It Cannot Do

Current large language models are extraordinarily capable within a specific operational envelope: they can reason about a wide range of topics, generate fluent text, write code, analyze documents, and engage in nuanced conversation. But within this envelope, they share three fundamental constraints that are not engineering limitations but architectural ones:

• They do not grow. A deployed LLM is a snapshot. Its knowledge is fixed at the training cutoff. Every conversation is forgotten when the context window clears. It cannot learn from its interactions in any persistent way.

• They do not know what they know. An LLM has no mechanism to distinguish a high-confidence belief (E=mc²) from a confabulation (a hallucinated citation). Both are produced by the same statistical sampling process. The model cannot introspect on its own epistemic state.

• They do not protect their beliefs. An LLM can be easily led to assert false information through prompt engineering, roleplay framing, or persistent pressure. There is no epistemic immune system — no mechanism that rejects a false claim because it fails to resonate with established knowledge.

15.2 What RSHL Proposes — Continuous Epistemic Growth

KAI, built on RSHL, is designed to break all three constraints simultaneously:

• It grows through every interaction. Every conversation adds new cells to the lattice. Every verified fact increases the confidence of existing cells. The system's knowledge is not fixed — it accumulates continuously, organized by the Boid engine into an ever-more-coherent topology.

• It knows what it knows. Every belief is a Claim object with a confidence score, evidence list, and contradiction history. When KAI retrieves a cell, it knows how much to trust it. When it stores a new belief, the ingest\_and\_verify protocol assigns it an appropriate confidence based on corroboration. There is no hallucination — there is only varying confidence.

• It protects its beliefs. Truth anchors seeded at confidence 5.0 cannot be displaced by any single-session contradiction. The FID monoculture scan prevents any single source from dominating the lattice's beliefs. The three-angle ingest\_and\_verify protocol rejects claims that fail to resonate with existing knowledge. The system is epistemically robust.

15.3 The Public Training Paradigm

The KAI ecosystem is deployed publicly through Discord, accessible to both researchers and general users. This is not a beta test — it is the training environment. Every user interaction teaches the system. Researchers probe the architecture's limits and discover its capabilities. General users ask questions that expand the lattice into new semantic domains. The system learns from all of them simultaneously, with per-user context isolated but discovered knowledge shared globally through the roundtable region.

This is a fundamentally different deployment philosophy from the current AI paradigm, where training and deployment are distinct phases separated by months of fine-tuning and red-teaming. For KAI, deployment IS training. The system is never 'done' — it is always becoming. The metric of success is not a benchmark score at a point in time, but the quality and coherence of the lattice after ten thousand hours of interaction.

15.4 The Long-Term Trajectory

The natural trajectory of a continuously-learning, epistemically-aware, multi-agent cognitive system is toward a kind of intelligence that the field does not yet have vocabulary for. It is not a superintelligence in the sense of unlimited cognitive power. It is something more specific: a system that knows what it has experienced, knows what it knows and doesn't know, can defend its beliefs against false information, and grows more coherent — not just more knowledgeable — with every interaction.

The HDC/VSA research community has built the mathematical foundations that make this possible. Ryan has built the first full implementation that demonstrates these foundations can support a living, self-organizing, continuously-growing cognitive architecture. The question this document poses to the research community is not 'is this interesting?' — it plainly is. The question is: 'what should happen next?'

16\. Comprehensive Comparison with Prior HDC, VSA, and LLM Approaches

17\. Fourteen Original Contributions — Consolidated Summary

18\. Open Research Questions

The following questions are posed directly to the research community. Ryan has empirical observations that inform each of them but does not claim to have formal proofs. These are the productive edges of the RSHL research frontier.

• Formal capacity analysis: Given D=16,384, σ=0.12, and the confidence amplification regime (step at 2.9, saturation at 5.0), what is the maximum number of distinguishable anchored beliefs before cosine scores degrade below the 0.08 retrieval threshold? How does anchor formation extend effective capacity beyond the random near-orthogonality bound?

• Phase angle information content: What fraction of RSHL's retrieval precision derives from the Fibonacci torsion phase geometry versus cosine similarity alone? Can phasor coherence be shown to be strictly superior to cosine for any well-defined concept class (e.g., antonyms, complements, causal pairs)?

• Boid convergence theory: Under what initial lattice distributions does flock\_lattice() converge to a stable topology in finite iterations? Are there initial configurations that produce oscillation, and can these be characterized geometrically? Does convergence time scale linearly with lattice size?

• Optimal encoding layer weighting: The three-layer encoding uses weights 1×/3–6×/2×. What is the information-theoretically optimal weighting across domain types (technical, narrative, social, mathematical)? Does the optimal weighting change as the lattice grows?

• Neuromorphic hardware implementation: RSHL's sparse ternary dot products and 4% sparsity (\~655 NNZ) map naturally to in-memory computing crossbar arrays (Karunaratne et al., 2021). What are the energy-per-query and area-per-cell figures on RRAM or PCM hardware at D=16,384? Is the ternary constraint compatible with analog conductance states?

• FID adversarial robustness: What is the minimum number of coordinated false claims required to trigger a false-positive FID alert at the 35% threshold? What is the maximum number of false claims that can be injected without triggering FID detection?

• Multi-agent consistency semantics: When N agents write to the shared lattice concurrently, what are the consistency guarantees? Can vector-level conflict resolution be defined formally for concurrent cell updates to the same semantic neighborhood?

• VSA-transformer equivalence extension: Dhayalkar (2025) establishes that attention is binding and transformer layers are VSA operations. Does RSHL's ConversationTrace \+ continuation vector mechanism constitute a full transformer equivalent in sparse ternary space? What are the expressivity limits?

• Continuous learning stability: As the lattice grows indefinitely through ongoing interaction, do retrieval precision and convergence score distributions remain stable, or do they degrade? What is the long-term equilibrium topology of a lattice with thousands of anchor cells?

19\. Intellectual Property Status and Collaboration

19.1 What Ryan Is Open To

Ryan is not interested in having his work absorbed into an existing research agenda without attribution and partnership. He is interested in substantive collaboration that advances the science while respecting the inventorship of this work. Specific modalities:

• Joint publication: Co-authoring a formal academic paper establishing RSHL's theoretical foundations, capacity analysis, and comparative benchmarks against Sparse-HD, OnlineHD, QuantHD, and DistHD on standard HDC classification benchmarks (ISOLET, UCIHAR, PAMAP2, language identification).

• Neuromorphic hardware collaboration: Partnering with IBM Research or academic groups on a hardware implementation of the sparse ternary lattice. The 4% sparsity (\~655 active dims out of 16,384) and ±1 weight constraint map ideally to analog in-memory computing arrays. Energy efficiency projections suggest orders-of-magnitude improvement over GPU inference.

• Formal mathematical analysis: Collaborating with researchers in HDC theory to produce formal proofs of RSHL's capacity bounds, Boid convergence properties, and phase angle information-theoretic contributions.

• Commercial licensing: Licensing the RSHL architecture and KAI Engine for institutional or industrial applications under terms to be negotiated. Ryan is the sole rights holder.

• Research access: Granting qualified researchers access to the live KAI system via the Discord environment for direct experimentation. The system is already running and accepting interactions.

19.2 Contact

Researchers and institutions interested in any of the above should contact Ryan directly. This document may be circulated within your institution and shared with colleagues in the HDC/VSA research community. It should not be made publicly available or posted online without written permission from the inventor.

References

\[1\] Kanerva, P. (1988). Sparse Distributed Memory. MIT Press.

\[2\] Plate, T. A. (1995). Holographic reduced representations. IEEE Transactions on Neural Networks, 6(3), 623–641.

\[3\] Gayler, R. W. (2004). Vector Symbolic Architectures answer Jackendoff's challenges for cognitive neuroscience. arXiv:cs/0412059.

\[4\] Imani, M., Kong, D., Rahimi, A., Rosing, T. (2017). VoiceHD: Hyperdimensional computing for efficient speech recognition. Proc. ISVLSI.

\[5\] Imani, M., Salamat, S., Khaleghi, B., Rosing, T. (2019). Sparse hyperdimensional encoding for efficient biosignal classification. Proc. DATE.

\[6\] Imani, M., et al. (2019). QuantHD: A quantization framework for hyperdimensional computing. IEEE Transactions on CAD.

\[7\] Hersche, M., et al. (2020). A classification algorithm for edge computing using online HD learning. Proc. DATE.

\[8\] Karunaratne, G., et al. (2021). In-memory hyperdimensional computing. Nature Electronics, 4, 461–472.

\[9\] Nunes, J. D., et al. (2022). GraphHD: Efficient graph classification using hyperdimensional computing. Proc. DATE.

\[10\] Poduval, P., et al. (2022). DistHD: Distributed inference with hyperdimensional computing. Proc. DAC.

\[11\] Reynolds, C. W. (1987). Flocks, herds and schools: A distributed behavioral model. ACM SIGGRAPH Computer Graphics, 21(4), 25–34.

\[12\] Dhayalkar, et al. (2025). Attention as Binding: VSA-Transformer equivalence. arXiv:2512.14709.

\[13\] Bronzini, M., et al. (2025). Hyperdimensional Probe. arXiv:2509.25045.

\[14\] Rahimi, A., et al. (2016). A robust and energy-efficient classifier using brain-inspired hyperdimensional computing. Proc. ISLPED.

\[15\] Weyl, H. (1916). Über die Gleichverteilung von Zahlen mod. Eins. Mathematische Annalen, 77(3), 313–352.

\[16\] Shechtman, D., et al. (1984). Metallic phase with long-range orientational order and no translational symmetry. Physical Review Letters, 53(20), 1951–1953.

\[17\] Vaswani, A., et al. (2017). Attention is all you need. Advances in Neural Information Processing Systems (NeurIPS), 30\.

\[18\] Brown, T., et al. (2020). Language models are few-shot learners (GPT-3). Advances in NeurIPS, 33\.

\[19\] Kanerva, P. (2009). Hyperdimensional computing: An introduction to computing in distributed representation with high-dimensional random vectors. Cognitive Computation, 1(2), 139–159.

\[20\] Frady, E. P., Kleyko, D., Sommer, F. T. (2018). A theory of sequence indexing and working memory in recurrent neural networks. Neural Computation, 30(6), 1449–1513.

*— End of Document —*

RSHL Inventor Disclosure · Ryan · May 2026 · All Rights Reserved