Skip to content
Genius Cognitive System
MNN::Transformer
Initializing search
Genius Cognitive System
Architecture
Architecture
1. Executive Summary
1. Executive Summary
Developer interfaces and commercialization boundary
2. System Objectives
2.1. Primary Goals
2.2. Secondary Goals
2.3. Cognitive Architecture and Component Roles
2.3. Cognitive Architecture and Component Roles
2.3.1 Executive Controller
2.3.2 Execution Integrity System
2.3.3 GAML as the Cognitive Knowledge Layer
2.3.4 Bridge Blocks and Context Assembly
2.3.5 Semantic Core and Specialist Cognition
2.3.6 Verification, Arbitration, and Synthesis
2.3.7 Learning and Distillation Feedback
2.3.8 Relationship to the GNUS Swarm
3 System Architecture Overview
4 GNUS Component Mapping
4.1 Compute Layer
4.1 Compute Layer
4.1.1 SGFP4 Design
4.2 Distributed Layer
4.2 Distributed Layer
4.2.1 Layered Cognitive Stack
4.3 Security Layer
4.3 Security Layer
4.3.1 Core Architectural Distinction
5 Model Architecture
5.1 Semantic Core
5.1 Semantic Core
5.1.1 Base Model
5.1.2 Quantization
5.2 Expert Models (EMs) and Specialist Modules
5.2 Expert Models (EMs) and Specialist Modules
5.2.1 Expert Model taxonomy
5.2.2 Expert Language Models (ELMs)
5.2.3 Expert Judgment Models (EJMs)
5.2.4 Diffusion-Backed Expert Processing
5.2.5 Role-Based Experts
5.2.6 Domain-Specific Experts
5.2.7 Private Expert Models
5.2.7.1 Expert Identity, Runtime Instance, and Memory Scope
5.2.8 Expert Invocation Patterns
5.2.9 Model Processing and Execution Backends
5.2.10 Legacy MVP Specialists
6 Router Design
6.1 Router and Planner Responsibilities
6.2 Initial MVP Router
6.3 Future Router Evolution
7 Reputation-Based Consensus System
7.1 Reputation Data Model
7.2 Reputation Update Formula
7.2 Reputation Update Formula
7.2.1 Accuracy / Quality Component
7.2.2 Latency Component
7.2.3 Consistency Component
7.2.4 Safety and Policy Component
7.2.5 Final Update
7.3 Weighted Consensus Algorithm
7.4 Consensus Engine Architecture (Protocol Layer)
7.4 Consensus Engine Architecture (Protocol Layer)
7.4.1 Consensus Design Principles
7.4.2 Swarm Execution Flow
7.4.3 Consensus Message Types
7.4.4 Liveness Model
7.4.5 Byzantine Tolerance
7.4.6 Reputation-Gated Participation
7.4.7 Genesis Anchor Model
8.4 GNUS Agentic Memory Layer (GAML v1)
8.4 GNUS Agentic Memory Layer (GAML v1)
8.4.1 Purpose
8.4.2 Architectural Position
8.4.3 Memory Object Model
8.4.4 Cognitive Asset Model
8.4.4.1 Memory Stage, Class, Scope, and State
8.4.4.2 Working, Short-Term, and Long-Term Memory
8.4.4.3 Activation and Retention Decay
8.4.4.4 Consolidation, Major Events, and Forgetting
8.4.4.5 Mistakes, Lessons, and Promotion
8.4.4.6 Compartmentalized Long-Term Memory and Sparse Overlays
8.4.4.7 Bounded Associative Activation
8.4.5 Ingestion Pipeline
8.4.6 Agentic Retrieval Mechanism
8.4.7 Surprise-Gated Writes
8.4.8 Memory as Support for Experts
8.4.9 Swarm Memory Consensus
8.4.10 Replication and Convergence
8.4.11 Private Memory, Ownership, and Derived Assets
8.4.11.1 Privacy Scopes
8.4.11.2 Ownership and Authorization
8.4.11.3 Inference Boundaries
8.4.11.4 Encryption and Key References
8.4.11.5 Derived-Asset Inheritance
8.4.11.6 Training and Adaptation
8.4.11.7 Local Personal Data
8.4.12 Performance & Overhead Impact
8.4.13 Strategic Impact
8 Grounding and Retrieval
8.1 Grokipedia Role
8.2 Retrieval Pipeline
8.3 Validation Layer
8.4 Private Knowledge Grounding
8.5 Grounding Modes
8.6 Grounding as an Expert Role
8.7 Why Retrieval Is Not Enough by Itself
8.7 Why Retrieval Is Not Enough by Itself
8.7.1 Extended Grounding Memory
9. Execution and Performance - Execution Modes and Performance Targets
9.1 Mode 1 — Single Node
9.2 Mode 2 — ELM-Assisted Mode
9.3 Mode 3 — Swarm Mode
9.4 Mode 4 — Agent Mode
9.5 Execution Strategy Principles
10 Performance Targets
11 Execution Roadmap
11.1 Phase 1 — Semantic Core Foundations
11.2 Phase 2 — Experts + Router / Planner
11.3 Phase 3 — Reputation, Memory, and Consensus
11.4 Phase 4 — Grounding, Private Customization, Secure Agent Path, and Benchmarks
12 Risk Analysis
13 Future Compatibility
14 Strategic Positioning
15 AI Safety Philosophy
15.1 Safety Architecture Model
15.1 Safety Architecture Model
15.1.1 Layer 1 — Node-Level Enforcement (Authoritative)
15.1.2 Layer 2 — Reputation-Based Enforcement
15.1.3 Layer 3 — Client-Side Preference Filtering
15.1.4 Layer 4 — Tool Intermediary Enforcement
15.2 Safety Profile Declaration
15.3 No GeoIP Enforcement
15.4 Grounding Safety Integration
15.5 Safety in Swarm Mode
15.6 Safety-Aware Expert Patterns
15.7 Compliance & Liability Model
16 Distributed Swarm Thinking Context Architecture
16.1 Purpose
16.2 Why this section exists
16.3 Architectural intent
16.4 Core design principles
16.4 Core design principles
16.4.1 Structured collaborative reasoning over monolithic reasoning
16.4.2 Memory-guided context instead of brute-force long context
16.4.3 Inspectable swarm thinking
16.4.4 Reputation-aware specialization
16.4.5 Quantization-aware modularity
16.5 System overview
16.5 System overview
16.5.1 High-level flow
16.6 Thinking context model
16.6 Thinking context model
16.6.1 Definition
16.6.2 Why this matters
16.7 Memory and context construction
16.7 Memory and context construction
16.7.1 Bridge Blocks
16.7.2 Fact store
16.7.3 Profile layer
16.7.4 Retrieval flow
16.8 Specialist taxonomy
16.8 Specialist taxonomy
16.8.1 Role specialists
16.8.2 Domain specialists
16.9 Recommended evolution from current specialists
16.9 Recommended evolution from current specialists
16.9.1 Current state
16.9.2 Recommended near-term state
16.9.3 Recommended medium-term state
16.10 Routing model
16.10 Routing model
16.10.1 MVP routing
16.10.2 Future learned routing
16.11 Execution patterns
16.11 Execution patterns
16.11.1 Core-only response
16.11.2 Sequential specialist chain
16.11.3 Distributed swarm execution
16.11.4 Streaming draft with delayed refinement
16.12 Thinking trace schema
16.12 Thinking trace schema
16.12.1 Example trace sections
16.13 Relation to consensus and reputation
16.13 Relation to consensus and reputation
16.13.1 Current score types
16.13.2 Recommended future score types
16.14 Interaction with SGFP4, Turbo Quant, and Sparse-V
16.14 Interaction with SGFP4, Turbo Quant, and Sparse-V
16.14.1 Semantic Core
16.14.2 Small specialists
16.14.3 Verifier and router models
16.14.4 Sparse-V implications
16.14.5 Open quantization questions
16.15 Adapter and distillation implications
16.15 Adapter and distillation implications
16.15.1 Specialist implementation and training identity
16.15.2 Distillation targets by role
16.15.3 Swarm trace to EJM training-shard conversion
16.16 Summary
17. Context Lifecycle, Caching, and Governance
17.1 Purpose
17.2 Normative Language
17.3 Scope
17.4 Non-Goals
17.5 Architectural Position
17.6 Core Design Principles
17.6 Core Design Principles
17.6.1 Smallest Sufficient Context
17.6.2 Off-Window by Default for Large Material
17.6.3 Stable Before Volatile
17.6.4 Deterministic Construction
17.6.5 Specialists as Lossy Filters
17.6.6 Compaction at Semantic Boundaries
17.6.7 Hierarchical Budgets
17.6.8 Explainability Without Raw Hidden Reasoning
17.6.9 Privacy Never Widens Through Optimization
17.7 Canonical Context Packet Model
17.8 Stable-Prefix Contract
17.8 Stable-Prefix Contract
17.8.1 Permitted Stable Material
17.8.2 Prohibited Volatile Material
17.8.3 Canonicalization
17.9 Prefix Cache Identity and Compatibility
17.9 Prefix Cache Identity and Compatibility
17.9.1 EJM Shared-State Cache Isolation
17.10 Cache Lifecycle and Invalidation
17.11 Context Selection and Hierarchical Token Allocation
17.11 Context Selection and Hierarchical Token Allocation
17.11.1 Budget Order
17.11.2 Stage Budgets
17.11.3 Selection Priority
17.11.4 Duplicate and Supersession Filtering
17.12 Off-Window Content and Artifact References
17.13 Expert Isolation and the ExpertDigest Contract
17.14 Capability Schema Loading
17.15 Context Compaction Lifecycle
17.15 Context Compaction Lifecycle
17.15.1 Compaction Modes
17.15.2 Preservation Requirements
17.15.3 Compaction Record
17.15.4 Hierarchical Bridge Blocks
17.15.5 Compaction Quality Check
17.16 Context Inspector
17.17 Routing and Scheduling Integration
17.18 Verification and Early Exit
17.19 Security, Privacy, and Trust Requirements
17.19 Security, Privacy, and Trust Requirements
17.19.1 Cache Isolation
17.19.2 No Secret Material in Model Prefixes
17.19.3 Revocation and Deletion
17.19.4 Cache Poisoning and Context Injection
17.19.5 Trust Separation
17.20 Observability and Evaluation
17.21 Failure Modes and Required Fallbacks
17.21 Failure Modes and Required Fallbacks
17.21.1 Over-Compaction
17.21.2 Prefix Churn
17.21.3 Stale Compatible Cache
17.21.4 Retrieval Omission
17.21.5 Expert Digest Loss
17.21.6 Cross-Scope Leakage
17.21.7 Budget Starvation
17.21.8 Cache-Centric Routing Regression
17.22 Canonical Interfaces and Data Contracts
17.22 Canonical Interfaces and Data Contracts
17.22.1 ContextInclusionDecision
17.22.2 PrefixCacheEvent
17.22.3 CompactionRecord
17.23 Initial Implementation Requirements
17.24 Acceptance Criteria
17.25 Rollout Plan
17.25 Rollout Plan
17.25.1 Phase One — Instrumentation
17.25.2 Phase Two — Deterministic Context Compilation
17.25.3 Phase Three — Prefix Cache Management
17.25.4 Phase Four — Expert Isolation and Lazy Schemas
17.25.5 Phase Five — Compaction and Inspection
17.25.6 Phase Six — Adaptive Optimization
17.26 Design Principle
17.27 Non-Normative Motivation
18 Secure Agent Architecture for the GNUS.ai Decentralized Cognitive System
18.1 Product Technical Design Specification
18.1 Product Technical Design Specification
18.1.1 Goals and Success Criteria
18.1.2 System Overview
18.1.3 Core Components
18.1.4 End-to-End Data Flows
18.1.5 Interfaces and Data Contracts
18.1.6 Reliability, Fault Tolerance, and Quality Control
18.1.7 MVP Implementation Mapping
18.1.8 Metrics and Observability
18.1.9 Open Decisions for Next Iteration
18.1.10 Implementation Notes and Recommendations
18.1.11 Hand-off Instructions for Next Engineer or LLM
18.1.12 Summary
19. EGGROLL Swarm Retraining Architecture
19.1 Purpose
19.2 Architectural Position
19.3 Why EGGROLL Fits GNUS.ai
19.4 Design Principles
19.4 Design Principles
19.4.1 Locality First
19.4.2 Deterministic Reconstruction over Tensor Shipment
19.4.3 Compact Fitness over Gradient Exchange
19.4.4 Adapter-Oriented Evolution
19.4.5 Reputation-Gated Promotion
19.4.6 Hierarchical Swarm Aggregation
19.4.7 Calibrated Judgments
19.5 Relationship to Adapters and Expert Execution
19.6 Core Training Primitive
19.6 Core Training Primitive
19.6.1 EJM Distillation Storage and Sharding
19.7 GNUS Processing Room Mapping
19.8 Beehives and Locality-Aware Sub-Swarms
19.9 Deterministic Perturbation Reconstruction
19.10 Worker Execution Model
19.11 Fitness Packet Design
19.12 Aggregation Model
19.13 Reputation and Validation Extensions
19.14 Embedded Retraining Loop
19.14 Embedded Retraining Loop
19.14.1 Normal Inference Path
19.14.2 Learning Event Creation
19.14.3 Retraining Conversion
19.14.4 Artifact Publication
19.15 Best Initial Retraining Targets
19.15 Best Initial Retraining Targets
19.15.1 Numeric Specialist / Math Verifier
19.15.2 Router / Planner Specialist
19.15.3 Formatter / Schema Specialist
19.15.4 Grounding Specialist
19.15.5 Code Specialist
19.15.6 Expert Judgment Models (EJMs)
19.16 Safety and Governance Constraints
19.17 Constraints and Non-Goals
19.18 Rollout Plan
19.18 Rollout Plan
19.18.1 Phase 1 — Single-Machine Proof
19.18.2 Phase 2 — Local Beehive
19.18.3 Phase 3 — GNUS Processing Room Integration
19.18.4 Phase 4 — Reputation and Redundancy
19.18.5 Phase 5 — Hierarchical Swarm Aggregation
19.19 Strategic Positioning
19.20 Summary
20. Targeted Retraining and Hierarchical Critical Thinking Specialists
20.1 Overview
20.2 Targeted Retraining
20.2 Targeted Retraining
20.2.1 Key Properties
20.3 EGGROLL-Based Optimization
20.3 EGGROLL-Based Optimization
20.3.1 Why EGGROLL
20.3.2 Optimization Targets
20.3.3 Reward Signals
20.4 Hierarchical Critical Thinking Specialists (HCTS)
20.4 Hierarchical Critical Thinking Specialists (HCTS)
20.4.1 Hierarchical Structure
20.4.2 Persistent Contextual Memory Overlays
20.5 Functional Responsibilities
20.6 Bias-Aware Reasoning
20.7 Cognitive Resistance Layer
20.7 Cognitive Resistance Layer
20.7.1 Modes
20.7.2 Adaptive Friction Triggers
20.8 Integration with Cognitive Twin
20.9 Continuous Learning Loop
20.10 System Outcome
20.11 Summary
21. Data-Driven Epistemic Arbitration and Cognitive OS Extensions
21.1 Purpose
21.2 Why this section exists
21.3 Architectural intent
21.4 Core design principles
21.4 Core design principles
21.4.1 Arbitration is a first-class cognitive function
21.4.2 Epistemic frameworks are modular and swappable
21.4.3 Framework logic should be data-driven
21.4.4 The Requestor Node is the correct control point
21.4.5 Inspectable reasoning should not depend on raw chain-of-thought exposure
21.4.6 Plugins should remain extremely small
21.5 Relationship to the existing Genius architecture
21.5 Relationship to the existing Genius architecture
21.5.1 Relation to the Semantic Core
21.5.2 Relation to ELMs and experts
21.5.3 Relation to consensus
21.5.4 Relation to grounding
21.5.5 Relation to GAML
21.5.6 Relation to HCTS
21.6 Requestor Node as Epistemic Arbiter
21.6 Requestor Node as Epistemic Arbiter
21.6.1 Current role of the Requestor Node
21.6.2 Extended role
21.6.3 Why this is the right place
21.6.4 Cognitive OS implication
21.7 Why GQHSM is the correct runtime
21.7 Why GQHSM is the correct runtime
21.7.1 Problem shape
21.7.2 GQHSM as the execution substrate
21.7.3 Why not hardcode the frameworks directly
21.7.4 Determinism and inspectability
21.7.5 Selective EJM/JDM judgment callbacks
21.8 Native implementation model: C++, MNN, and separation of concerns
21.8 Native implementation model: C++, MNN, and separation of concerns
21.8.1 Execution stack
21.8.2 Separation of concerns
21.8.3 Why this is efficient
21.8.4 Why this fits mobile and desktop deployment
21.9 Supported epistemic framework families
21.9 Supported epistemic framework families
21.9.1 Sanskrit epistemology
21.9.2 Kripke and modal reasoning
21.9.3 Hybrid frameworks
21.9.4 Future frameworks
21.10 Sanskrit epistemology as a practical arbitration model
21.10 Sanskrit epistemology as a practical arbitration model
21.10.1 Why Sanskrit reasoning is useful here
21.10.2 Mapping the phases into Genius
21.10.3 Why this is better than simple weighted merge
21.11 Kripke modal arbitration in practical system terms
21.11 Kripke modal arbitration in practical system terms
21.11.1 Why modal reasoning belongs here
21.11.2 World construction
21.11.3 Accessibility and survivability
21.11.4 Fixed-point resolution
21.12 Hybrid arbitration strategies
21.12 Hybrid arbitration strategies
21.12.1 Sequential hybrid
21.12.2 Parallel hybrid
21.12.3 Why hybridization matters
21.13 GQHSM machine structure
21.13 GQHSM machine structure
21.13.1 Structural requirements
21.13.2 Representative machine outline
21.13.3 Sanskrit branch outline
21.13.4 Kripke branch outline
21.13.5 Hybrid branch outline
21.14 JSON-defined machine configuration
21.14 JSON-defined machine configuration
21.14.1 Why configuration matters
21.14.2 Example machine definition
21.14.3 Why this matters
21.15 Generic callback model
21.15 Generic callback model
21.15.1 Context and lifecycle callbacks
21.15.2 Core reasoning callbacks
21.15.3 Guard callbacks
21.15.4 Why generic callbacks matter
21.16 Plugin architecture
21.16 Plugin architecture
21.16.1 Why plugins are the right shape
21.16.2 What a plugin does
21.16.3 Stable plugin ABI
21.16.4 Example plugin shape
21.16.5 Operational advantages
21.17 Future WASM extension path
21.17 Future WASM extension path
21.17.1 Why WASM is attractive later
21.17.2 Why not require it first
21.17.3 Forward compatibility
21.18 Epistemic context model
21.18 Epistemic context model
21.18.1 Required inputs
21.18.2 Why this context matters
21.19 Example plugin and loader behavior
21.19 Example plugin and loader behavior
21.19.1 Example registration flow
21.19.2 Example loader shape
21.20 Output model and thinking trace
21.20 Output model and thinking trace
21.20.1 Example trace artifact
21.20.2 Why this is important
21.21 Integration with memory writeback and retraining
21.21 Integration with memory writeback and retraining
21.21.1 Memory writeback
21.21.2 Retraining implications
21.22 Strategic implications
21.22 Strategic implications
21.22.1 Why this matters competitively
21.22.2 Why this matters architecturally
21.23 Risks and open questions
21.24 Summary
22 SGFP4 Adaptive Quantization Format
22.1 Design Goals
22.2 Macroblocks (Tiling)
22.3 v1 Profile — Container Layout
22.3 v1 Profile — Container Layout
22.3.1 Alignment and Flags-in-Offsets
22.4 Affine Parameters (Scale + Bias)
22.5 Per-Block Mode Flags
22.6 Quantization Modes and Payload Packing
22.6 Quantization Modes and Payload Packing
22.6.1 FP4_AFFINE (MODE = 0)
22.6.2 T158_AFFINE (MODE = 1)
22.7 Encoding (Non-Normative)
22.8 v2 Profile — Quadtree-Adaptive Stream
22.8 v2 Profile — Quadtree-Adaptive Stream
22.8.1 File Framing
22.8.2 Record Layout
22.8.3 v2 Decode Procedure
22.9 GPU Decode Procedure (v1)
22.10 Conformance and Bit-Exact Replay
22.11 Cross-Referencing
23 Objective Memory and Verified Transition Graph (VTG)
23.1 Purpose
23.2 Architectural Position
23.3 Why this layer exists
23.4 Objective vs. Subjective Cognition
23.5 Verified Transition Graph
23.6 State Identity
23.7 Transition Edge Model
23.8 Candidate Frontier
23.8 Candidate Frontier
23.8.1 Negative Transition Evidence and Scoped Activation
23.9 Relationship to GAML
23.10 Relationship to Swarm Thinking Context
23.11 Relationship to Router and Planner
23.12 Relationship to Semantic Core and Expert Models
23.13 Relationship to Epistemic Arbitration
23.14 Relationship to HCTS and Subjective Preference
23.15 Relationship to EGGROLL
23.15 Relationship to EGGROLL
23.15.1 Relationship to EJM Decision Corpora and Distillation Shards
23.16 Storage and Distribution Model
23.17 Update Semantics
23.18 Security and Poisoning Resistance
23.19 Privacy Model
23.20 Performance Model
23.21 Initial Implementation Path
23.21 Initial Implementation Path
23.21.1 Phase 1 — Instrumentation Only
23.21.2 Phase 2 — Local VTG Prototype
23.21.3 Phase 3 — Verified Candidate Frontier
23.21.4 Phase 4 — Tenant-Private VTG
23.21.5 Phase 5 — Swarm Replication
23.21.6 Phase 6 — EGGROLL Optimization
23.22 Non-Goals
23.23 Strategic Impact
23.24 Summary
24 Speculative Decoding and VTG Candidate Scheduling
24.1 Purpose
24.2 Operating Envelope
24.3 Why this layer exists
24.4 Core Components
24.5 Micro-Speculation Backend Classes
24.6 Confidence-Scheduled Prefix Retention
24.7 VTG as the Primary Swarm Advantage
24.8 Micro-Diffusion Block Drafting
24.9 Tiny Causal Tree Drafting
24.10 Frozen Micro-MTP as the First Neural Target
24.11 Role-Specific Speculation Policy
24.12 Node Capability Advertisement
24.13 Swarm Outcome Events
24.14 Integration with EGGROLL
24.15 Initial Implementation Plan
24.15 Initial Implementation Plan
24.15.1 Phase 1 — Instrumentation
24.15.2 Phase 2 — VTG Lookup + Rule Drafter
24.15.3 Phase 3 — Frozen Micro-MTP Head
24.15.4 Phase 4 — Tiny Causal Tree Head
24.15.5 Phase 5 — Micro-Diffusion Block Drafter
24.15.6 Phase 6 — Swarm Optimization
24.16 Scope Boundaries
24.17 Design Principle
25 Frozen Micro-MTP and VTG Edge Inference
25.1 Purpose
25.2 Operating Envelope
25.3 Why this matters
25.4 Core Design Principle
25.5 Micro-MTP Budget
25.6 Relationship to VTG
25.7 Candidate Record
25.8 Best Initial Targets
25.9 Local Verification Requirements
25.10 Node Capability Advertisement
25.11 Relationship to Micro-Diffusion and Tiny Tree Drafting
25.12 Relationship to EGGROLL
25.13 Initial Implementation Path
25.13 Initial Implementation Path
25.13.1 Phase 1 — Measurement
25.13.2 Phase 2 — Formatter / Schema Micro-MTP
25.13.3 Phase 3 — Code Specialist Micro-MTP
25.13.4 Phase 4 — Router Policy
25.13.5 Phase 5 — Swarm Learning
25.14 Summary
26 OpenAI-Compatible API Router and GCS Job Queue Architecture
26.1 Product Technical Design Specification
26.2 Background and Current Queue Context
26.3 Goals
26.3 Goals
26.3.1 Primary Goals
26.3.2 Developer Experience Goals
26.4 Non-Goals for MVP
26.5 Core Design Principle
26.6 Job Type Split
26.6 Job Type Split
26.6.1 Existing Job Type: Processing Chunk Job
26.6.2 New Job Type: API Request Job
26.6.3 Why This Split Matters
26.7 Architecture Overview
26.8 Components
26.8 Components
26.8.1 Cloudflare Edge
26.8.2 GCS API Router
26.8.3 GCS Gateway Node
26.8.4 Online GCS Worker Nodes
26.8.5 Router / Planner Node
26.8.6 Aggregator Node
26.9 Pub/Sub Channels
26.9 Pub/Sub Channels
26.9.1 Capability Registration Channels
26.9.2 API Job Channels
26.9.3 Processing Chunk Channels
26.9.4 Result, Stream, and Claim Channels
26.10 Node Registration
26.10 Node Registration
26.10.1 Registration Envelope
26.10.2 Heartbeat Rules
26.11 API Request Job Envelope
26.11 API Request Job Envelope
26.11.1 Required Fields
26.11.2 Payload Storage Modes
26.11.3 Routing Modes
26.12 Claim, Lock, and Lease Semantics
26.12 Claim, Lock, and Lease Semantics
26.12.1 First Valid Claim MVP
26.12.2 Claim Envelope
26.12.3 Lease Rules
26.13 API Job Lifecycle
26.13 API Job Lifecycle
26.13.1 States
26.13.2 Lifecycle Flow
26.13.3 Cancellation Flow
26.14 Native Processing Jobs and ELM Work Items
26.14 Native Processing Jobs and ELM Work Items
26.14.1 One Funded ELM Job for Multiple ELMs
26.14.2 Parent API Correlation and Completion
26.15 OpenAI-Compatible API Surface
26.15 OpenAI-Compatible API Surface
26.15.1 /v1/models
26.15.2 /v1/chat/completions
26.15.3 /v1/embeddings
26.15.4 GNUS Extension Object
26.16 Streaming Proxy Requirements
26.16 Streaming Proxy Requirements
26.16.1 Streaming Proxy Responsibilities
26.16.2 Recommended Streaming Headers
26.16.3 First-Token Behavior
26.16.4 Internal-to-External Stream Bridge
26.16.5 Internal Stream Chunk
26.16.6 External OpenAI-Compatible Chunk
26.16.7 Backpressure and Slow Clients
26.16.8 Disconnect and Cancellation Semantics
26.16.9 Requeue During Streaming
26.16.10 Pre-Stream vs Post-Stream Errors
26.16.11 Stream Ordering and Replay Protection
26.16.12 Cloudflare and Runtime Boundary
26.16.13 Streaming Test Matrix
26.16.14 Streaming Acceptance Criteria
26.17 Result Envelope
26.17 Result Envelope
26.17.1 Final Result
26.17.2 Error Result
26.18 Queue Fairness and Democratized Pickup
26.18 Queue Fairness and Democratized Pickup
26.18.1 MVP Policy
26.18.2 Later Policies
26.18.3 Fairness State
26.19 Requeue and Retry
26.19 Requeue and Retry
26.19.1 Requeue Reasons
26.19.2 Requeue Envelope
26.20 Security and Privacy
26.20 Security and Privacy
26.20.1 API Key Security
26.20.2 Job Signature Requirements
26.20.3 Prompt Privacy
26.20.4 Public Queue Leakage
26.21 Metering, Rewards, and Settlement
26.21 Metering, Rewards, and Settlement
26.21.1 Usage Record
26.21.2 Settlement Hooks
26.21.3 Native GNUS escrow and owner-set ELM-hour funding (October 2026)
26.22 Data Model Changes
26.22 Data Model Changes
26.22.1 New Message Concepts
26.22.2 Existing Processing Messages Remain
26.22.3 Optional Unified Queue Wrapper
26.23 CRDT Keyspace Proposal
26.24 MVP Implementation Plan
26.24 MVP Implementation Plan
26.24.1 Phase 1: API Compatibility Shell
26.24.2 Phase 2: API Request Job Schema
26.24.3 Phase 3: GCS Gateway Bridge
26.24.4 Phase 4: Node Registration and Claim
26.24.5 Phase 5: Direct Worker Execution
26.24.6 Phase 6: Funded ELM Job Bridging and Optional Non-ELM Jobs
26.24.7 Phase 7: Metering and Settlement
26.24.8 Phase 8: Private and Hybrid Routing
26.25 Acceptance Criteria
26.26 Open Questions
26.27 Summary
27. Local Cognitive Second Brain Mode
27.1 Purpose
27.2 Architectural Position
27.3 Orchestration Layer Role
27.4 GAML Role
27.5 Local and Private Expert Model Role
27.6 Second Brain Agent Role
27.6 Second Brain Agent Role
27.6.1 Supporting Agents and Modules
27.7 EGGROLL Role
27.8 High-Level Flow
27.9 Local Data Sources
27.10 Structured Memory Objects
27.11 Memory Lifecycle
27.11 Memory Lifecycle
27.11.1 Observe
27.11.2 Extract
27.11.3 Normalize
27.11.4 Link
27.11.5 Score
27.11.6 Store
27.11.7 Retrieve
27.11.8 Reason
27.11.9 Verify
27.11.10 Write Back
27.11.11 Adapt
27.12 Context Packet Assembly
27.13 Human-Readable Memory Mirror
27.14 Privacy Modes
27.14 Privacy Modes
27.14.1 Local-Only Mode
27.14.2 Private Enterprise Mode
27.14.3 Hybrid Mode
27.14.4 Explicit Swarm Contribution Mode
27.15 Example Workflows
27.15 Example Workflows
27.15.1 Meeting Prep
27.15.2 Project Drift Detection
27.15.3 Personal Daily Brief
27.15.4 Private Expert Adaptation
27.16 Implementation Requirements
27.17 Design Principle
27.18 Summary
28. Forecast-Driven Cognition and Predictive Prefetching
28.1 Purpose
28.2 Inspiration and Scope
28.3 Biological and Conversational Motivation
28.4 Architectural Position
28.5 Anticipatory Cognition Engine (ACE)
28.6 Forecast Domains
28.6 Forecast Domains
28.6.1 Intent Forecasting
28.6.2 Semantic-Trajectory Forecasting
28.6.3 Memory Forecasting
28.6.4 Expert and Model Forecasting
28.6.5 Tool Forecasting
28.6.6 Network and Node Forecasting
28.6.7 Verification Forecasting
28.6.8 Response and Turn-Taking Forecasting
28.7 Cognitive Execution Scheduler (CES)
28.8 Confidence-Based Preparation Policy
28.9 Multi-Hypothesis Forecasting
28.10 Bidirectional Voice Communication
28.11 Personal Forecast Models
28.12 Anticipatory Distillation
28.13 Forecast Training Objectives
28.14 Distillation Data Generation
28.15 EGGROLL Integration
28.16 GAML Integration
28.17 Distributed GNUS Integration
28.18 Privacy, Safety, and User Control
28.19 Failure Modes
28.19 Failure Modes
28.19.1 Over-Prefetching
28.19.2 Premature Semantic Commitment
28.19.3 Cache Pollution
28.19.4 Privacy Leakage
28.19.5 Feedback Collapse
28.19.6 Incorrect Turn-Taking
28.20 Observability and Evaluation
28.21 Initial Implementation Phases
28.21 Initial Implementation Phases
28.21.1 Phase One: Local Voice and Text Forecasting
28.21.2 Phase Two: Personal Forecast Models
28.21.3 Phase Three: Anticipatory Distillation
28.21.4 Phase Four: Distributed Predictive Scheduling
28.21.5 Phase Five: Future Multimodal Evidence
28.22 Implementation Requirements
28.23 Design Principle
29. Execution Integrity System (EIS)
29.1 Purpose
29.2 Motivation and Threat Model
29.2 Motivation and Threat Model
29.2.1 Why zk is out of scope at the GCS layer
29.2.2 The lazy-node / model-substitution attack
29.2.3 Security economics
29.2.4 Design principles
29.3 Execution Contracts
29.3 Execution Contracts
29.3.1 Definition
29.3.2 Contract pinning
29.4 Determinism Classes
29.4 Determinism Classes
29.4.1 Definition
29.4.2 Class A - Reference-integer semantics
29.4.3 Class B - Bounded-drift semantics
29.4.4 Class C - Non-deterministic paths
29.4.5 Constraints on kernel authors
29.5 Checkpoint-Band Matching
29.5 Checkpoint-Band Matching
29.5.1 Checkpoints, not per-op comparison
29.5.2 Comparison domain
29.5.3 Registration invariant: the band forgives hardware, not weights
29.5.4 Calibration and drift monitoring
29.5.5 EJM readout/head substitution
29.6 Teacher-Forced Spot-Check Protocol
29.6 Teacher-Forced Spot-Check Protocol
29.6.1 The autoregressive divergence problem
29.6.2 Teacher-forced replay
29.6.3 Cost profile
29.6.4 Sampling-consistency check
29.6.5 Scheduling and economics
29.7 Interaction with Semantic Consensus
29.8 Interaction with GAML and Cognitive Assets
29.9 Open Items
29.10 Cross-References
30. GCS Capability System
30.1 Purpose
30.2 Architectural Position
30.3 Core Concepts
30.3 Core Concepts
30.3.1 Capability
30.3.2 Connector
30.3.3 Capability Provider
30.3.4 Capability Contract
30.4 Canonical Capability Contract
30.5 Discovery and Contract Translation
30.6 MCP Connector Adapter
30.7 Connector Categories
30.7 Connector Categories
30.7.1 Local Personal Data
30.7.2 Developer and Engineering Systems
30.7.3 Enterprise and Productivity Systems
30.7.4 Data and Knowledge Systems
30.7.5 Finance, Commerce, and Blockchain
30.7.6 Device and Operating-System Services
30.7.7 GNUS and GCS Services
30.7.8 Optional External Inference Providers
30.8 Local Capability Execution
30.9 Capability Routing and Reputation
30.10 Credential and Secret Handling
30.11 Tool Intermediary Integration
30.12 GAML Integration
30.13 Connector Lifecycle
30.14 Initial Implementation Requirements
30.15 Design Principle
31. Agent and Module Development Inventory
31.1 Purpose
31.2 Inventory Conventions
31.2 Inventory Conventions
31.2.1 Component Classes
31.2.2 Required Definition for Each Component
31.2.3 Trust Tiers
31.3 Reference Runtime Topology
31.4 Executive Control and Orchestration Inventory
31.4.1 Ingress Service
31.4.2 Session and Identity Context Service
31.4.3 Executive Controller
31.4.4 Intent and Risk Classifier
31.4.5 Router
31.4.6 Planner
31.4.7 Execution Graph Compiler
31.4.8 Policy Evaluator
31.4.9 Budget and Constraint Manager
31.4.10 Task Coordinator
31.4.11 Scheduler and Dispatcher
31.4.12 Execution Trace Recorder
31.5 Semantic Core and Expert Execution Inventory
31.5.1 Semantic Core Runtime
31.5.2 Model Runtime Abstraction
31.5.3 Model and Artifact Registry
31.5.4 Expert Registry
31.5.5 Role-Based Expert Models
31.5.6 Domain-Specific Expert Models
31.5.7 Expert Output Packager
31.5.8 Private and Local Expert Model Manager
31.5.9 Expert Evaluation Harness
31.6 Forecast-Driven Cognition Inventory
31.6.1 Anticipatory Cognition Engine
31.6.2 Personal Forecast Model
31.6.3 Forecast Hypothesis Manager
31.6.4 Prefetch Planner
31.6.5 Cognitive Execution Scheduler
31.6.6 Speculative State Store
31.6.7 Streaming Voice Predictor
31.6.8 Forecast Outcome Evaluator
31.6.9 Anticipatory Distillation Exporter
31.7 GAML and Cognitive Asset Inventory
31.7.1 Cognitive Asset Schema Registry
31.7.2 GAML Memory API
31.7.3 Local Memory Store
31.7.4 Content-Addressed Artifact Store
31.7.5 Graph Relationship Service
31.7.6 Metadata, Semantic, and Graph Index Manager
31.7.7 Source Observer Agents
31.7.8 Ingestion and Normalization Service
31.7.9 Extraction Agents
31.7.10 Entity Resolution Service
31.7.11 Temporal and Supersession Resolver
31.7.12 Contradiction Detection Service
31.7.13 Provenance and Lineage Service
31.7.14 Memory Scoring and Write Gate
31.7.14 Memory Scoring and Write Gate
31.7.14.1 Memory Lifecycle and Consolidation Manager
31.7.14.2 Scoped Activation State Service
31.7.14.3 Cognitive Scope and Memory Overlay Service
31.7.14.4 Mistake and Lesson Processor
31.7.15 Private-Memory Authorization Service
31.7.16 Private-Memory Encryption and Key Reference Service
31.7.17 Derived-Artifact Privacy Inheritance
31.7.18 Memory Governor
31.7.19 Retrieval Planner
31.7.20 Bridge Block Generator
31.7.21 Context Packet Builder
31.7.22 CRDT Replication and Convergence Service
31.7.23 Human-Readable Memory Mirror
31.7.24 Retention, Deletion, and Revocation Service
31.7.25 Memory-to-Learning Export Gate
31.8 Objective Memory and VTG Inventory
31.8.1 VTG State Canonicalizer
31.8.2 Transition Edge Store
31.8.3 Candidate Frontier Service
31.8.4 Transition Validation Collector
31.8.5 Transition Decay and Compatibility Manager
31.8.6 VTG Trace Integration
31.9 Grounding, Verification, Arbitration, and Synthesis Inventory
31.9.1 Claim Extraction Service
31.9.2 Public Grounding Service
31.9.3 Private Grounding Service
31.9.4 Evidence Packager
31.9.5 Specialized Verifiers
31.9.6 Confidence and Uncertainty Calibrator
31.9.7 Epistemic Framework Registry
31.9.8 Arbitration Runtime
31.9.9 Reputation-Weighted Consensus Coordinator
31.9.10 Synthesis Service
31.9.11 Final Formatter
31.10 Reputation and Consensus Infrastructure
31.10.1 Reputation Store
31.10.2 Reputation Update Engine
31.10.3 Sybil and Abuse Resistance
31.10.4 Reputation Replication
31.11 Execution Integrity System Inventory
31.11.1 Execution Contract Builder
31.11.2 Kernel and Runtime Registry
31.11.3 Determinism Certification Service
31.11.4 Checkpoint Calibration Service
31.11.5 Execution Claim Collector
31.11.6 Spot-Check Scheduler
31.11.7 Teacher-Forced Execution Checker
31.11.8 Fraud Verdict Service
31.11.9 EIS Evidence Store
31.11.10 Future Hardware Attestation Adapter
31.12 GCS Capability System and Connector Inventory
31.12.1 Capability Registry
31.12.2 Connector Registry
31.12.3 Connector Discovery Manager
31.12.4 Capability Translation Agent
31.12.5 Capability Contract Validator
31.12.6 Capability Contract Signer and Version Manager
31.12.7 Policy Binder
31.12.8 Credential Broker
31.12.9 Capability Router
31.12.10 Provider Health and Reputation Monitor
31.12.11 Connector Drift Detector
31.12.12 GCS Capability Server
31.12.13 Initial Connector Families
31.12.13 Initial Connector Families
Local Personal Data
Development Systems
Data and Knowledge Systems
Enterprise Systems
Web and Research
Finance and Commerce
Device and Operating-System Capabilities
GNUS-Native Services
Optional External Inference Providers
31.12.14 Initial Connector Priority
31.13 Secure Tool Intermediary Inventory
31.13.1 Tool Proposal Intake
31.13.2 Capability Enforcement Engine
31.13.3 Dry-Run Engine
31.13.4 Sandbox Orchestrator
31.13.5 Output Sanitizer
31.13.6 Prompt-Injection and Trap Detector
31.13.7 Human Approval Service
31.13.8 Execution Broker
31.13.9 Tool Attestation Signer
31.13.10 Tool Result Normalizer
31.13.11 Memory Writeback Gate for Tool Results
31.14 Local Cognitive Second Brain Inventory
31.14.1 Second Brain Agent
31.14.2 Local Source Observer
31.14.3 Personal Entity and Project Resolver
31.14.4 Commitment and Deadline Tracker
31.14.5 Meeting Preparation Service
31.14.6 Daily Brief Service
31.14.7 Project Drift Detector
31.14.8 Local Email Capability
31.14.9 Personal Vault Mirror
31.14.10 Private Device Synchronization
31.14.11 User Correction and Confirmation Service
31.15 OpenAI-Compatible API and GCS Job Queue Inventory
31.15.1 OpenAI-Compatible API Router
31.15.2 Authentication and Rate-Limit Service
31.15.3 API Request Job Builder
31.15.4 GCS Gateway Node
31.15.5 API Request Queue
31.15.6 Processing Chunk Queue
31.15.7 Node Capability Registration
31.15.8 Streaming Channel Manager
31.15.9 Result Aggregator
31.15.10 Usage, Metering, and Settlement Adapter
31.16 Distributed Infrastructure Inventory
31.16.1 libp2p Node Service
31.16.2 Pub/Sub Topic Manager
31.16.3 Distributed Task Queue
31.16.4 Node Discovery and Capability Advertisement
31.16.5 Distributed Artifact Manager
31.16.6 CRDT State Manager
31.16.7 gRPC Service Interface Layer
31.16.8 Node Health Monitor
31.16.9 Node Identity and Signing
31.17 EGGROLL and Adaptive Learning Inventory
31.17.1 Learning Signal Collector
31.17.2 Privacy and Curation Gate
31.17.3 Distillation and Decision Corpus Builder
31.17.3.1 Distillation View Compiler
31.17.3.2 EJM State and Branch Shard Builder
31.17.3.3 Judgment Calibration and Robustness Evaluator
31.17.4 Fitness Evaluator
31.17.5 Perturbation and Training Coordinator
31.17.6 Promotion Gate
31.17.7 Local and Tenant Adaptation Manager
31.17.8 Learning Artifact Registry
31.18 Security, Privacy, and Governance Inventory
31.18.1 Identity and Authorization Service
31.18.2 Privacy Policy Engine
31.18.3 Key Management Service
31.18.4 Secrets Service
31.18.5 Sandbox Profile Registry
31.18.6 Audit Event Store
31.18.7 Consent and Approval Record Service
31.18.8 Threat Detection Service
31.19 User and Operator Interface Inventory
31.19.1 Cognitive Client Interface
31.19.2 Capability Approval Interface
31.19.3 Memory Inspector
31.19.4 Connector Administration Interface
31.19.5 Node and Swarm Operations Interface
31.19.6 Evaluation and Benchmark Interface
31.20 Observability and Operational Services
31.20.1 Canonical Event Bus
31.20.2 Trace Service
31.20.3 Metrics Service
31.20.4 Logging Service
31.20.5 Alerting Service
31.20.6 Configuration and Feature-Flag Service
31.21 Canonical Interface and Schema Inventory
31.22 Deployment Profiles
31.22.1 Single-Device Local Profile
31.22.2 Personal Multi-Device Profile
31.22.3 Private Enterprise Profile
31.22.4 Public GNUS Swarm Profile
31.22.5 Hybrid Profile
31.22.6 Dedicated Tool Intermediary Profile
31.22.7 Dedicated Verification and EIS Profile
31.23 Logical Workstreams and Package Boundaries
31.24 Recommended Delivery Sequence
31.24.1 Milestone 0 — Contracts and Security Boundaries
31.24.2 Milestone 1 — Local Cognitive Baseline
31.24.3 Milestone 2 — Private Memory and Second Brain
31.24.4 Milestone 3 — Capability System and Secure Tool Path
31.24.5 Milestone 4 — Verification and Distributed Execution
31.24.6 Milestone 5 — Execution Integrity
31.24.7 Milestone 6 — VTG, Forecasting, and Learning
31.24.8 Milestone 7 — Ecosystem Hardening
31.25 Validation and Test Inventory
31.25.1 Unit Tests
31.25.2 Integration Tests
31.25.3 Security Tests
31.25.4 Privacy Tests
31.25.5 Reliability Tests
31.25.6 Performance Tests
31.25.7 Quality Evaluation
31.26 Component Definition of Done
31.27 Summary
GNUS.ai and GCS: Architecture at a Glance
The idea in two paragraphs
Boundaries readers should not confuse
Real implementation versus the plan
Document Index and Source Authority
Start here
Normative or current references
Material that must be labeled historical or superseded
Investor-reader rule
Implementation Status and Evidence Register
Status vocabulary
Versioned evidence table
Conflicts and gaps requiring owner action
Four-chain EVM token supply: direct read recipe
Evidence standard for upgrades
Performance Benchmarks: Evidence and Protocol
What must be measured
Required baseline matrix
Reproducible result record
Evidence status
GNUS.ai Technical Diligence FAQ
GNUS.ai Technical Diligence FAQ
Is GNUS.ai just an inexpensive GPU-cloud provider?
Is mainnet finished or launched?
Does everyone vote with reputation points?
Is the processing burn 10%, and do burns guarantee token appreciation?
Is a GNUS divided into a billion Minions?
Does a matching GPU result hash prove a correct answer?
What about the old 'dummy for now' verifier seed?
Does GCS run a complete expert/memory/evaluation pipeline today?
Who schedules the distributed ELMs?
What is the verified price?
Have the node, bridges and GCS been audited?
Are token supply, returns, and tax status confirmed by architecture documents?
Cognitive Evolution Coordination
Purpose
Architectural Position
Qualified Cognitive Event
Qualified Cognitive Event
Schema and Signature Contract
Governance Resolution
Coordination Flow
Shared Contracts
Shared Contracts
Event Contract
EJM Decision Replay and Distillation Lineage
Replay Contract
Promotion Contract
Runtime and Repository Ownership
Observability
Summary
Developer API and Distributed Compute Bridge
The two developer entry points
Business models enabled by the same infrastructure
Implemented native job funding: GNUS-priced escrow
Owner-resolved ELM-job rate (designed, not yet deployed)
Delivery and validation
GNUS-NEO-SWARM Source
GNUS-NEO-SWARM Source
Classes
Classes
Args
FlutterWindow
GeniusElmFFI
PipelineTest
Win32Window
Win32Window
None
Point
Size
WindowClassRegistrar
sgns
sgns
neoswarm
neoswarm
ChainStep
CognitiveAsset
ELMContext
ExecutionChain
InferenceResponse
KnowledgeFact
MemoryContext
NodeOutput
NodeReputation
PromptFeatures
RouteDecision
Task
api
api
ApiServer
ApiServer
None
Config
Config
None
ElmEntry
core
core
InferenceEngine
MNNInferenceEngine
MNNInferenceEngine
None
Config
SGProcessingBridge
SGProcessingBridge
None
Config
TensorInterpreter
Tokenizer
elm
elm
DomainELM
ELMChainBuilder
ELMChainBuilder
None
Config
GroundingELM
IELM
RoleELM
SpecialistAdapter
ToolSupportELM
knowledge
knowledge
ContextInjection
ContextInjection
None
Config
FactValidation
FactValidation
None
ValidationResult
KnowledgeRetrieval
KnowledgeRetrieval
None
Config
Impl
Impl
None
FactEntry
network
network
GeniusNodeConfig
P2PNode
P2PNode
None
Config
Impl
Impl
None
GossipSubs
ResultAggregation
ResultAggregation
None
Config
SGClient
SGClient
None
Config
Impl
SGJobSubmitter
SGResultCollector
SGResultCollector
None
Impl
SGResultCollectorConfig
reputation
reputation
ReputationCRDT
ReputationScoring
ReputationScoring
None
Config
ReputationStorage
ReputationStorage
None
Impl
WeightedConsensus
WeightedConsensus
None
Config
router
router
IRouter
PromptAnalyzer
RuleBasedRouter
RuleBasedRouter
None
Config
security
security
MessageSigning
NodeIdentity
NodeIdentity
None
Impl
specialists
specialists
GrammarSpecialist
ISpecialist
MathSpecialist
SymbolicFallback
SymbolicFallback
None
Parser
storage
storage
GcsGlobalDb
GcsGlobalDb
None
Config
test
test
GcsGlobalDbTest
Files
Files
GNUS-NEO-SWARM
GNUS-NEO-SWARM
neoswarm_ffi
neoswarm_ffi
example
example
ios
ios
Runner
Runner
Runner/Runner-Bridging-Header.h
linux
linux
runner
runner
runner/my_application.h
windows
windows
runner
runner
runner/flutter_window.cpp
runner/flutter_window.h
runner/neoswarm_ffi/example/windows/runner/main.cpp
runner/resource.h
runner/utils.cpp
runner/utils.h
runner/win32_window.cpp
runner/win32_window.h
ios
ios
Classes
Classes
Classes/ios/Classes/flutter_slm_bridge.c
macos
macos
Classes
Classes
Classes/macos/Classes/flutter_slm_bridge.c
src
src
src/flutter_slm_bridge.h
src/os_defines.h
src/src/flutter_slm_bridge.c
src
src
api
api
api/api_server.cpp
api/api_server.hpp
common
common
common/error.cpp
common/error.hpp
common/logging.hpp
common/types.hpp
core
core
engine
engine
engine/inference_engine.hpp
engine/mnn_inference_engine.cpp
engine/mnn_inference_engine.hpp
sgprocessing
sgprocessing
sgprocessing/sg_processing_bridge.cpp
sgprocessing/sg_processing_bridge.hpp
sgprocessing/tensor_interpreter.cpp
sgprocessing/tensor_interpreter.hpp
tokenizer
tokenizer
tokenizer/tokenizer.hpp
elm
elm
elm/domain_elm.cpp
elm/domain_elm.hpp
elm/elm_chain_builder.cpp
elm/elm_chain_builder.hpp
elm/grounding_elm.cpp
elm/grounding_elm.hpp
elm/i_elm.hpp
elm/role_elm.cpp
elm/role_elm.hpp
elm/specialist_adapter.cpp
elm/specialist_adapter.hpp
elm/tool_support_elm.cpp
elm/tool_support_elm.hpp
src/genius_elm_chat_c.cpp
src/genius_elm_chat_completions.cpp
src/genius_elm_chat_completions.h
knowledge
knowledge
knowledge/context_injection.cpp
knowledge/context_injection.hpp
knowledge/fact_validation.cpp
knowledge/fact_validation.hpp
knowledge/knowledge_retrieval.cpp
knowledge/knowledge_retrieval.hpp
memory
memory
memory/memory_init.cpp
network
network
network/p2p_node.cpp
network/p2p_node.hpp
network/result_aggregation.cpp
network/result_aggregation.hpp
sg_client
sg_client
sg_client/sg_job_submitter.cpp
sg_client/sg_job_submitter.hpp
sg_client/sg_result_collector.cpp
sg_client/sg_result_collector.hpp
sg_client/super_genius_client.cpp
sg_client/super_genius_client.hpp
reputation
reputation
reputation/node_reputation.hpp
reputation/reputation_crdt.cpp
reputation/reputation_crdt.hpp
reputation/reputation_scoring.cpp
reputation/reputation_scoring.hpp
reputation/reputation_storage.cpp
reputation/reputation_storage.hpp
reputation/weighted_consensus.cpp
reputation/weighted_consensus.hpp
router
router
router/i_router.hpp
router/prompt_analyzer.cpp
router/prompt_analyzer.hpp
router/rule_based_router.cpp
router/rule_based_router.hpp
security
security
security/message_signing.cpp
security/message_signing.hpp
security/node_identity.cpp
security/node_identity.hpp
specialists
specialists
specialists/grammar_specialist.cpp
specialists/grammar_specialist.hpp
specialists/i_specialist.hpp
specialists/math_specialist.cpp
specialists/math_specialist.hpp
specialists/symbolic_fallback.cpp
specialists/symbolic_fallback.hpp
src/src/main.cpp
storage
storage
storage/gcs_global_db.cpp
storage/gcs_global_db.hpp
test
test
benchmark
benchmark
benchmark/bench_mnn_llm.cpp
benchmark/os_memory.hpp
common
common
common/test_types.cpp
elm
elm
elm/test_chain_builder.cpp
elm/test_elm.cpp
ffi
ffi
ffi/test_genius_elm_ffi.cpp
integration
integration
integration/test_pipeline.cpp
integration/test_sg_connectivity.cpp
integration/test_sgprocessing_pipeline.cpp
knowledge
knowledge
knowledge/test_context_injection.cpp
knowledge/test_fact_validation.cpp
knowledge/test_knowledge_retrieval.cpp
network
network
network/test_network.cpp
network/test_sg_client.cpp
reputation
reputation
reputation/test_reputation.cpp
router
router
router/test_router.cpp
security
security
security/test_message_signing.cpp
security/test_node_identity.cpp
specialists
specialists
specialists/test_grammar_specialist.cpp
specialists/test_math_specialist.cpp
specialists/test_symbolic_fallback.cpp
storage
storage
storage/test_gcs_global_db.cpp
Namespaces
Namespaces
MNN
MNN
None
Transformer
boost
boost
None
asio
libp2p
libp2p
None
basic
sgns
sgns
None
crdt
ipfs_lite
ipfs_lite
None
ipfs
ipfs
None
graphsync
ipfs_pubsub
neoswarm
neoswarm
None
api
core
elm
knowledge
memory
network
reputation
router
security
specialists
storage
storage
None
test
sgns::neoswarm::api
sgns::neoswarm::core
sgns::neoswarm::core
sgns::neoswarm::core
sgns::neoswarm::elm
sgns::neoswarm::elm
sgns::neoswarm::elm
sgns::neoswarm::elm
sgns::neoswarm::elm
sgns::neoswarm::elm
sgns::neoswarm::knowledge
sgns::neoswarm::knowledge
sgns::neoswarm::network
sgns::neoswarm::network
sgns::neoswarm::network
sgns::neoswarm::network
sgns::neoswarm::network
sgns::neoswarm::reputation
sgns::neoswarm::reputation
sgns::neoswarm::reputation
sgns::neoswarm::reputation
sgns::neoswarm::router
sgns::neoswarm::router
sgns::neoswarm::security
sgns::neoswarm::security
sgns::neoswarm::specialists
sgns::neoswarm::specialists
sgns::neoswarm::storage
std
testing
Python (gnus-poc)
Python (gnus-poc)
Classes
Classes
config
config
loader
loader
ConfigLoader
ConfigValidationError
distill
distill
backends
backends
anthropic_backend
anthropic_backend
AnthropicBackend
base
base
TeacherBackend
openai_backend
openai_backend
OpenAIBackend
cascade
cascade
CascadeResult
TeacherCascade
convergence
convergence
ConvergenceConfig
ConvergenceState
ConvergenceTracker
distillation
distillation
Distiller
sweep_analyzer
sweep_analyzer
SweepAnalyzer
synthetic
synthetic
SyntheticDataGenerator
teacher
teacher
TeacherClient
_ResponseWrapper
teacher_errors
teacher_errors
BackendNotFoundError
BudgetExceededError
CircuitBreakerOpenError
SyntheticDataError
TeacherConfigError
eval
eval
benchmark_config
benchmark_config
ConfigError
benchmark_mlx_model
benchmark_mlx_model
MLXBenchmarkModel
benchmark_runner
benchmark_runner
BenchmarkRunner
benchmarker
benchmarker
Benchmarker
MissingBaselineError
evaluator
evaluator
SpecialistEvaluator
metric_store
metric_store
EvalMetrics
MetricStore
threshold_adapter
threshold_adapter
ThresholdAdapter
ThresholdRecommendation
pipeline
pipeline
checkpoint
checkpoint
CheckpointValidator
StageValidationResult
runner
runner
PipelineRunner
StageResult
quantize
quantize
fp4_exporter
fp4_exporter
FP4Exporter
laplacian
laplacian
LaplacianWeightedError
manifest
manifest
ManifestBuilder
quadtree
quadtree
QuadtreeEncoder
sgfp4_decoder
sgfp4_decoder
SGFP4FormatError
sgfp4_format
sgfp4_format
CodeMode
Layout
router
router
plan
plan
ExecutionPlan
rules
rules
RouterConfigError
RuleEngine
state_machine
state_machine
RouterStateMachine
training
training
adapter_validator
adapter_validator
AdapterValidator
ValidationResult
config
config
TrainingConfig
tracker
tracker
ExperimentTracker
Files
Files
GNUS-NEO-SWARM
GNUS-NEO-SWARM
gnus-poc
gnus-poc
config
config
config/config/init.py
config/loader.py
data
data
scripts
scripts
scripts/analyze_common_pile.py
scripts/data/scripts/init.py
scripts/extract_source_niches.py
scripts/prepare_datasets.py
distill
distill
backends
backends
backends/anthropic_backend.py
backends/base.py
backends/distill/backends/init.py
backends/openai_backend.py
distill/cascade.py
distill/convergence.py
distill/distill/init.py
distill/distillation.py
distill/sweep_analyzer.py
distill/synthetic.py
distill/teacher.py
distill/teacher_errors.py
eval
eval
eval/benchmark_config.py
eval/benchmark_fingerprint.py
eval/benchmark_mlx_model.py
eval/benchmark_repair.py
eval/benchmark_runner.py
eval/benchmark_tasks.py
eval/benchmark_trends.py
eval/benchmarker.py
eval/eval/init.py
eval/evaluator.py
eval/metric_store.py
eval/threshold_adapter.py
pipeline
pipeline
pipeline/checkpoint.py
pipeline/pipeline/init.py
pipeline/runner.py
quantize
quantize
quantize/fp4_exporter.py
quantize/laplacian.py
quantize/manifest.py
quantize/quadtree.py
quantize/quantize/init.py
quantize/sgfp4_decoder.py
quantize/sgfp4_format.py
router
router
router/plan.py
router/router/init.py
router/rules.py
router/state_machine.py
training
training
training/adapter_validator.py
training/config.py
training/dedup.py
training/memory.py
training/tokenizer_utils.py
training/tracker.py
training/train_specialists.py
training/train_specialists_mlx.py
training/training/init.py
Namespaces
Namespaces
config
config
None
loader
distill
distill
None
backends
backends
None
anthropic_backend
base
openai_backend
cascade
convergence
distillation
sweep_analyzer
synthetic
teacher
teacher_errors
eval
eval
None
benchmark_config
benchmark_fingerprint
benchmark_mlx_model
benchmark_repair
benchmark_runner
benchmark_tasks
benchmark_trends
benchmarker
evaluator
metric_store
threshold_adapter
pipeline
pipeline
None
checkpoint
runner
quantize
quantize
None
fp4_exporter
laplacian
manifest
quadtree
sgfp4_decoder
sgfp4_format
router
router
None
plan
rules
state_machine
scripts
scripts
None
analyze_common_pile
extract_source_niches
prepare_datasets
std
training
training
None
adapter_validator
config
dedup
memory
tokenizer_utils
tracker
train_specialists
train_specialists_mlx
MNN::Transformer
¶
Updated on 2026-10-09 at 03:59:42 +0000
Back to top