MACLLM Chain

The World's First Blockchain That Thinks

Whitepaper v2.0 — March 2026

Surasak Khankasikam (ocorpt)
MACLLM Project

Table of Contents

1. Introduction
2. The Problem with Centralized AI
3. Architecture
4. MACLLM Brain — Probability Matrix Engine
5. ORIN Chain — The Body
6. Proof of Cognition
7. Knowledge Protection
8. Matrix Sharding
9. Dual Token Economy
10. AI Safety
11. Scaling Strategy
12. Developer Experience
13. MVP Results
14. Roadmap
15. Conclusion

1. Introduction

The current AI landscape is dominated by centralized services where a handful of corporations control the world's most powerful language models. Users have no ownership over the intelligence they help create, no way to verify how their data is used, and no provenance for AI-generated outputs.

MACLLM Chain addresses these fundamental issues by creating a decentralized AI network where:

• Anyone can use AI without centralized gatekeeping
• Intelligence is protected through mathematical properties, not legal agreements
• Every query is recorded on-chain with full provenance
• Contributors (compute, knowledge, capital) are automatically compensated
• The network gets smarter as more participants join

2. The Problem with Centralized AI

2.1 Data Sovereignty

When users interact with centralized AI services, their queries, conversations, and data reside on corporate servers. Users cannot verify whether their data is used for training, shared with third parties, or truly deleted upon request.

2.2 No Provenance

AI outputs lack traceability. There is no way to verify what information sources were used, what confidence level the system had, or which computational resources contributed to the answer.

2.3 Monopolistic Economics

Users who teach AI through feedback and corrections generate value that accrues entirely to the service provider. Knowledge contributors receive no ongoing compensation for the intelligence they help create.

3. Architecture

MACLLM Chain consists of two complementary layers:

    User Node          Cognitive Node          Knowledge Node
        |                    |                       |
        v                    v                       v
  +----------------------------------------------------------+
  |                    ORIN Chain (Body)                      |
  |  Ledger | Consensus | mCoin | Provenance | Smart Rules   |
  +----------------------------------------------------------+
        |                    |                       |
        v                    v                       v
  +----------------------------------------------------------+
  |                   MACLLM Brain (Mind)                     |
  |  6-Stage Pipeline | 64-Axis Composite Embedding (65,536d) | Matrix Sharding  |
  |  Triple Semantic Encoder (Koopman + PPMI + Neural) | GenerativeEngine  |
  +----------------------------------------------------------+

3.1 Three Node Types

3.2 Six-Step Query Flow

1. ORIN: Record query transaction, check mCoin balance, lock payment
2. Brain: Matrix route query, split task across Cognitive Nodes
3. Nodes: Perform blind computation on partial matrix shards
4. Brain: Reassemble partial results, generate answer
5. ORIN: Record answer transaction, distribute mCoin, store provenance hash
6. Return: Deliver answer with provenance proof to user

4. MACLLM Brain — Probability Matrix Engine

Unlike transformer-based LLMs that require billions of parameters and massive GPU clusters, MACLLM uses a probability matrix architecture that runs on commodity hardware (Apple Mac Mini M1, 16GB RAM).

4.1 64-Axis Composite Embedding (65,536d)

Every query is represented as a 64-axis composite embedding (64 × 1,024d = 65,536 dimensions) powered by a Triple Semantic Encoder (Koopman Eigenfunctions + PPMI+SVD + Mini Neural). The axes are organized across six core groups:

The full 64-axis space fits in a single CPU cache line as a 64×64 matrix = 32KB. The Triple Semantic Encoder produces a 1,024-dimensional vector per axis: Koopman Eigenfunctions (342d) capture temporal dynamics, PPMI+SVD (342d) encodes co-occurrence statistics from 100K documents, and a Mini Neural encoder (340d) provides discriminative triplet-trained representations.

4.2 Five-Level Knowledge Compression

Inspired by how the human brain compresses experience into principles rather than memorizing raw facts:

4.3 Key Modules

The brain comprises 84+ modules including: ThinkingPipeline (6-stage reasoning), CompositionEngine (16-module text generation), GenerativeEngine (creates text without LLM using ConceptGraph + PatternBank + FlowModel), Triple Semantic Encoder (Koopman + PPMI + Neural), Koopman Dynamics (predictive modeling), GeniusNetwork (multi-engine assembly), and SelfThinking (autonomous learning).

5. ORIN Chain — The Body

5.1 Consensus

ORIN uses BFT (Byzantine Fault Tolerant) consensus with BLS signature aggregation and VRF-based leader rotation. Finality is immediate within a shard (no forks). A minimum of 2/3+1 validators must sign each block.

5.2 Ledger

The ledger enforces a Fixed-Supply Invariant: the total supply of mCoin (1 billion) is verified at every checkpoint. No minting is allowed after genesis. State-first validation means transactions are checked against current state without replaying history.

5.3 Provenance

Every query/answer pair is recorded as a provenance entry containing: query hash, answer hash, contributing cognitive nodes, knowledge sources, confidence score, latency, and axis snapshot. Raw data is never stored on-chain — only cryptographic hashes.

5.4 Five Core Functions

1. Ledger: Record mCoin/MACLLM token balances
2. Identity: Genesis-Anchored Accounts with reputation and uptime
3. Rules: Immutable smart contracts (even the architect cannot modify)
4. Audit Trail: Every query hash on-chain for provenance
5. Consensus: Proof of Cognition + VRF verifiers

6. Proof of Cognition

Traditional consensus mechanisms waste energy (PoW) or favor the wealthy (PoS). Proof of Cognition validates that nodes can actually perform cognitive work.

6.1 Three Core Mechanisms

1. Hidden Test Query: The Brain inserts queries with known answers into normal traffic. Nodes don't know which queries are tests.
2. Cross-Validation: The same query is sent to multiple nodes. Results are compared — outliers are flagged.
3. Gradient Check: Verifies that a node has a functioning matrix (not returning random noise) by checking output structure against expected mathematical properties.

6.2 Scoring

Nodes must maintain a cognitive score of ≥60% to remain active. Reputation is updated using exponential moving average: rep = 0.9 × old + 0.1 × score. Failing nodes are automatically deactivated.

7. Knowledge Protection

The core intellectual property of MACLLM Chain is protected through four layers:

7.1 Anti-Frequency Analysis

Four countermeasures prevent statistical pattern recognition:

1. Matrix Rotation: Axis assignments change every N hours (+0ms, done on schedule)
2. Noise Injection: Deterministic random noise per request (+0.01ms)
3. Axis Shuffling: Axis order changes per session/node (+0.05ms)
4. Multi-path Routing: Same request sent to different shards (+0ms, already parallel)

Total overhead: +0.06ms. Crucially, more usage makes the system harder to analyze, not easier.

8. Matrix Sharding

The probability matrix is split into N partitions (default: 3) with configurable overlap (default: 10%) for redundancy. Each Cognitive Node receives only its assigned shard and performs blind matrix multiplication without knowing what the values represent.

The Brain reassembles partial results by averaging overlapping regions. Rotation changes shard boundaries periodically, and noise is added before transmission to nodes.

9. Dual Token Economy

9.1 MACLLM Token (Equity)

• Fixed supply: 1,000,000,000 (no inflation)
• Governance rights + staking
• Price determined by market
• 10% of query fees distributed to holders

9.2 mCoin (Utility)

• Soft peg target: ~$0.01 USD
• Compute-backed (not reserve-backed)
• Used to pay for all AI services
• Stable pricing regardless of MACLLM token price

9.3 Revenue Distribution per Query

10. AI Safety

MACLLM Chain separates the unstoppable (ORIN chain for money/tokens) from the controllable (AI layer):

• Hierarchy: Architect > System Goals > Avoid Harm
• Behavioral Watchdog: Rule-based system separate from AI — cannot be manipulated by the AI
• Koopman Dynamics: Detects trajectory drift before problems manifest
• Emergency Pause: Architect can pause AI layer for 72 hours; community vote for permanent action
• DNA Lock: AI can improve within its Learning Layer but cannot modify its core architecture

11. Scaling Strategy

Following the principle "Compress Down, Not Scale Up" — like the human brain, which doesn't grow larger when it becomes smarter:

12. Developer Experience

ORIN Chain provides comprehensive developer tooling to lower the barrier to building on the network:

12.1 Python SDK

A zero-dependency Python client (orin/sdk.py) wraps the full REST API with typed dataclasses. Developers can query, transfer mCoin, check provenance, and monitor chain health in 3 lines of code. The SDK auto-detects local vs. public endpoints and supports API key authentication.

12.2 REST API & Authentication

The ORIN API (FastAPI) exposes 15+ endpoints for query flow, chain inspection, account management, and node monitoring. Read endpoints are public; write endpoints enforce per-IP rate limiting (30/min write, 120/min read). API keys use SHA-256 hashing with admin-only generation via localhost. Interactive documentation is available via Swagger UI at /docs.

12.3 Block Explorer

A client-side block explorer provides real-time visibility into blocks, transactions, provenance records, and account balances with search and auto-refresh capabilities.

12.4 SQLite Persistence

Chain state (blocks, accounts, provenance) persists across restarts using SQLite in WAL mode. The system auto-restores full in-memory state from disk on startup, ensuring zero data loss during upgrades or crashes.

12.5 Cross-Session User Memory

MACLLM Brain maintains per-user conversation memory across sessions. Topics, entities, and summaries are persisted to disk and injected as context when users return, enabling personalized long-term interactions without external databases.

13. MVP Results

As of March 2026, the working prototype demonstrates:

14. Roadmap

15. Conclusion

MACLLM Chain represents a fundamentally new approach to AI: one where intelligence is decentralized but protected, where every unit of energy performs actual cognitive work, and where contributors are automatically compensated through on-chain provenance. The working MVP demonstrates that this vision is technically feasible on commodity hardware, with a brain that scores 922/1000 (92.2%) on benchmarks and a chain that processes queries in sub-millisecond time with full provenance.

The system's unique combination of probability matrix architecture, Proof of Cognition consensus, multi-layer knowledge protection, and dual token economics creates a platform that becomes both more powerful and more secure as it grows — the opposite of traditional centralized AI services.

MACLLM Chain Whitepaper v2.0 — March 2026
Contact: [email protected] | macllm.ai
© 2026 MACLLM. All rights reserved.